Oracle EE

Recently, I was trying to setup TDE. Doing that I found out the Oracle provided documentation isn’t overly clear, and there is a way to do it in pre-Oracle 12, which is done using ‘alter system’ commands, and a new-ish way to do it in Oracle 12, using ‘administer key management’ commands. I am using version, so decided to use the ‘administer key management’ commands. This blogpost is about an exception which I see is encountered in the Januari 2017 (170117) PSU of the Oracle database, which is NOT happening in Oracle 12.2 (no PSU’s for Oracle 12.2 at the time of writing) and Oracle April 2016 and October 2016 PSU’s.

In order to test the wallet functionality for TDE, I used the following commands:

SQL> select status, wrl_parameter from v$encryption_wallet;


SQL> !mkdir /u01/app/oracle/admin/test/wallet

SQL> administer key management create keystore '/u01/app/oracle/admin/test/wallet' identified by "this_is_the_keystore_password";

keystore altered.

SQL> administer key management set keystore open identified by "this_is_the_keystore_password";

keystore altered.

SQL> administer key management set key identified by "this_is_the_keystore_password" with backup;
administer key management set key identified by "this_is_the_keystore_password" with backup
ERROR at line 1:
ORA-28374: typed master key not found in wallet

SQL> select status, wrl_parameter from v$encryption_wallet;


SQL> administer key management set keystore open identified by "this_is_the_keystore_password";

keystore altered.

SQL> select status, wrl_parameter from v$encryption_wallet;


Line 1-10: The DB_UNIQUE_NAME of the instance is ‘test’, and therefore the default wallet location is /u01/app/oracle/admin/test/wallet (ORACLE_BASE/admin/DB_UNIQUE_NAME/wallet). The wallet directory doesn’t exist by default, so I created it (line 10).
Line 12: Here the keystore/wallet is created with a password.
Line 16: After the wallet is created without auto-login, the wallet must be opened using the ‘set keystore open’ command.
Line 20: After the wallet has been created, it does not contain a master key. This is done using the ‘set key’ command. However, this throws an ORA-28374 error.
Line 26: After an error involving the wallet has occurred, the wallet closes.
Line 35: The wallet can simply be opened using the earlier used ‘set keystore open’ command.
Line 39: This is where the surprise is: after opening, the master key “magically” appeared (visible by the status ‘OPEN’, without a master key this would be ‘OPEN_NO_MASTER_KEY’).

I yet have to start creating encrypted table spaces. There might be more surprises, I can’t tell at this moment because I didn’t try it. However, once I discovered this oddity, I talked to my colleague Matt who gave me his own runbook for enabling TDE, which turned out to be the exact same list of commands as I compiled, however he did not encounter the ORA-28374 which I did. I tested the same sequence of commands on, (October 2016) and (April 2016) and there the ORA-28374 was not raised during execution of the ‘set key’ command.

Reading through My Oracle Support note Master Note For Transparent Data Encryption ( TDE ) (Doc ID 1228046.1), I found the following text:

All the versions after


As of If the key associated with the SYSTEM, SYSAUX or UNDO tablespaces is not present in the wallet you cannot associate a new master key with the database (i.e. you cannot activate that master key for the database) unless you set a hidden parameter :

SQL> administer key management use key ‘AUQukK/ZR0/iv26nuN9vIqcAAAAAAAAAAAAAAAAAAAAAAAAAAAAA’ identified by “welcome1” with backup;
administer key management use key ‘AUQukK/ZR0/iv26nuN9vIqcAAAAAAAAAAAAAAAAAAAAAAAAAAAAA’ identified by “welcome1” with backup
ERROR at line 1:
ORA-28374: typed master key not found in wallet

alter system set “_db_discard_lost_masterkey”=true;

SQL> administer key management use key ‘AUQukK/ZR0/iv26nuN9vIqcAAAAAAAAAAAAAAAAAAAAAAAAAAAAA’ identified by “welcome1” with backup;

The heading and first line read weird, the heading indicates the paragraph is about ‘all the versions after’ (which to me means 12.2), and the first line in the paragraph says ‘as of’, which very clearly says this is about version and higher. However, a little further it shows the exact error (ORA-28374) I encountered, and explains that if a current key is used in the data dictionary (mind data dictionary, not wallet), you must set “_db_discard_lost_masterkey” to true before you can create and use another master key for a wallet if you start over (wipe or move the wallet directory).

This makes sense to me now! I tried dropping and creating new wallets in my current 170117 PSU instance, and only tried creating an encryption wallet in a brand new freshly created instance. So if I would have EXACTLY done the same in the instances with the other PSU’s, which is repeatedly create and drop a wallet for TDE, I would have encountered the same ORA-28374 error. Well…I see this as a safety mechanism, be it not a very obvious one, not exuberant documented, and probably causing more grief than it would save if you run into the need the change the master key.


When sifting through a sql_trace file from Oracle version 12.2, I noticed a new wait event: ‘PGA memory operation’:

WAIT #0x7ff225353470: nam='PGA memory operation' ela= 16 p1=131072 p2=0 p3=0 obj#=484 tim=15648003957

The current documentation has no description for it. Let’s see what V$EVENT_NAME says:

SQL> select event#, name, parameter1, parameter2, parameter3, wait_class 
  2  from v$event_name where name = 'PGA memory operation';

------ ------------------------------------- ---------- ---------- ---------- ---------------
   524 PGA memory operation                                                   Other

Well, that doesn’t help…

Let’s look a bit deeper then, if Oracle provides no clue. Let’s start with the strace and sql_trace combination. For the test, I am doing a direct path full table scan on a table. Such a scan must allocate a buffer for the results (direct path reads do not go into the buffercache, table contents are scanned to the PGA and processed from there).

TS@fv122b2 > alter session set events 'sql_trace level 8';

Session altered.

Now use strace to look at the system calls in another session:

# strace -e write=all -e all -p 9426
Process 9426 attached

Now execute ‘select count(*) from t2’. The output is rather verbose, but the important bits are:

io_submit(140031772176384, 1, {{data:0x7f5ba941ffc0, pread, filedes:257, buf:0x7f5ba91cc000, nbytes:106496, offset:183590912}}) = 1
mmap(NULL, 2097152, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS|MAP_NORESERVE, -1, 0x4ee000) = 0x7f5ba8fbd000
mmap(0x7f5ba8fbd000, 1114112, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_FIXED|MAP_ANONYMOUS, -1, 0) = 0x7f5ba8fbd000
lseek(7, 0, SEEK_CUR)                   = 164639
write(7, "WAIT #0x7f5ba9596310: nam='PGA m"..., 112) = 112
 | 00000  57 41 49 54 20 23 30 78  37 66 35 62 61 39 35 39  WAIT #0x7f5ba959 |
 | 00010  36 33 31 30 3a 20 6e 61  6d 3d 27 50 47 41 20 6d  6310: nam='PGA m |
 | 00020  65 6d 6f 72 79 20 6f 70  65 72 61 74 69 6f 6e 27  emory operation' |
 | 00030  20 65 6c 61 3d 20 37 38  30 20 70 31 3d 32 30 39   ela= 780 p1=209 |
 | 00040  37 31 35 32 20 70 32 3d  31 31 31 34 31 31 32 20  7152 p2=1114112  |
 | 00050  70 33 3d 30 20 6f 62 6a  23 3d 32 32 38 33 33 20  p3=0 obj#=22833  |
 | 00060  74 69 6d 3d 31 39 35 31  37 30 32 30 35 36 36 0a  tim=19517020566. |
munmap(0x7f5ba8fbd000, 2097152)         = 0
munmap(0x7f5ba91bd000, 2097152)         = 0
mmap(0x7f5ba949d000, 65536, PROT_NONE, MAP_PRIVATE|MAP_FIXED|MAP_ANONYMOUS|MAP_NORESERVE, -1, 0x2ce000) = 0x7f5ba949d000
lseek(7, 0, SEEK_CUR)                   = 183409
write(7, "WAIT #0x7f5ba9596310: nam='PGA m"..., 100) = 100
 | 00000  57 41 49 54 20 23 30 78  37 66 35 62 61 39 35 39  WAIT #0x7f5ba959 |
 | 00010  36 33 31 30 3a 20 6e 61  6d 3d 27 50 47 41 20 6d  6310: nam='PGA m |
 | 00020  65 6d 6f 72 79 20 6f 70  65 72 61 74 69 6f 6e 27  emory operation' |
 | 00030  20 65 6c 61 3d 20 35 39  32 20 70 31 3d 30 20 70   ela= 592 p1=0 p |
 | 00040  32 3d 30 20 70 33 3d 30  20 6f 62 6a 23 3d 32 32  2=0 p3=0 obj#=22 |
 | 00050  38 33 33 20 74 69 6d 3d  31 39 35 32 30 36 33 33  833 tim=19520633 |
 | 00060  36 37 34 0a                                       674.             |

Okay, we can definitely say the mmap() and munmap() system calls seem to be related, which makes sense if you look a the name of the wait event. Let’s look a bit more specific using a systemtap script:

global wait_event_nr=524
probe begin {

probe process("/u01/app/oracle/product/").function("kskthbwt") {
	if ( pid() == target() && register("rdx") == wait_event_nr )
		printf("kskthbwt - %d\n", register("rdx"))
probe process("/u01/app/oracle/product/").function("kskthewt") {
	if ( pid() == target() && register("rsi") == wait_event_nr )
		printf("kskthewt - %d\n", register("rsi"))
probe syscall.mmap2 {
	if ( pid() == target() )
		printf(" mmap, addr %x, size %d, protection %d, flags %d, fd %i, offset %d ", u64_arg(1), u64_arg(2), int_arg(3), int_arg(4), s32_arg(5), u64_arg(6))
probe syscall.mmap2.return {
	if ( pid() == target() )
		printf("return value: %x\n", $return)
probe syscall.munmap {
	if ( pid() == target() )
		printf(" munmap, addr %x, size %d\n", u64_arg(1), u64_arg(2))

Short description of this systemtap script:
Lines 6-9: This probe is triggered once the function kskthbwt is called. This is one of the functions which are executed when the wait interface is called. The if function on line 7 checks if the process specified with -x with the systemtap executable is the process calling this function, and if the register rdx contains the wait event number. This way all other waits are discarded. If the wait event is equal to wait_event_nr, which is set to the wait event number 524, which is ‘PGA memory operation’, the printf() function prints kskthbwt and the wait event number. This is simply to indicate the wait has started.
Lines 10-13: This probe does exactly the same as the previous probe, except the function is kskthewt, which is one of the functions called when the ending of a wait event is triggered.
Line 14-17: This is a probe that is triggered when the mmap2() system call is called. Linux actually uses the second version of the mmap call. Any call to mmap() is silently executed as mmap2(). Inside the probe, the correct process is selected, and the next line simply prints “mmap” and the arguments of mmap, which I picked from the CPU registers. I do not print a newline.
Line 18-21: This is a return probe of the mmap2() system call. The function of this probe is to pick up the return code of the system call. For mmap2(), the return code is the address of the memory area mapped by the kernel for the mmap2() call.
Line 22-25: This is a probe on munmap() system call, which frees mmap’ed memory to the operating system.
Please mind there are no accolades following the if statements, which means the code executed when the if is true is one line following the if. Systemtap and C are not indention sensitive (like python), I indented for the sake of clarity.

I ran the above systemtap script against my user session and did a ‘select count(*) from t2’ again:

# stap -x 9426 mmap.stp
kskthbwt - 524
 mmap, addr 0, size 2097152, protection 3, flags 16418, fd -1, offset 750 return value: 7f5ba91bd000
 mmap, addr 7f5ba91bd000, size 1114112, protection 3, flags 50, fd -1, offset 0 return value: 7f5ba91bd000
kskthewt - 524
kskthbwt - 524
 mmap, addr 0, size 2097152, protection 3, flags 16418, fd -1, offset 1262 return value: 7f5ba8fbd000
 mmap, addr 7f5ba8fbd000, size 1114112, protection 3, flags 50, fd -1, offset 0 return value: 7f5ba8fbd000
kskthewt - 524
kskthbwt - 524
 munmap, addr 7f5ba8fbd000, size 2097152
 munmap, addr 7f5ba91bd000, size 2097152
kskthewt - 524

This makes it quite clear! The event ‘PGA memory operation’ is called when mmap() and munmap() are called. Which are calls to allocate and free memory for a process. The file descriptor (fd) value is set to -1, which means no file is mapped, but anonymous memory.

Another interesting thing is shown: first mmap is called with no address given, which makes the kernel pick a memory location. This memory location is then used for a second mmap call at the same memory address. The obvious question for this is: why mmap two times?

To answer that, we need to look at the flags of the two calls. Here is an example:

mmap(NULL, 2097152, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS|MAP_NORESERVE, -1, 0x4ee000) = 0x7f5ba8fbd000
mmap(0x7f5ba8fbd000, 1114112, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_FIXED|MAP_ANONYMOUS, -1, 0) = 0x7f5ba8fbd000

The first mmap call asks the kernel for a chunk of memory. PROT_READ and PROT_WRITE mean the memory should allow reading and writing. MAP_PRIVATE means it’s not public/shared, which is logical for Oracle PGA memory. MAP_ANONYMOUS means the memory allocation is not backed by a file, so just an allocation of contiguous memory. MAP_NORESERVE means no swap space is reserved for the allocation. This means this first mapping is essentially just a reservation of the memory range, no physical memory pages are allocated.

The next mmap call maps inside the memory allocated with the first mmap call. This seems strange at first. If you look closely at the flags, you see that MAP_NORESERVE is swapped for MAP_FIXED. The reason for this strategy to make it easier for the Oracle database to allocate the memory allocations inside a contiguous chunk of (virtual) memory.

The first mmap call allocates a contiguous (virtual) memory area, which is really only a reservation of a memory range. No memory is truly allocated, hence MAP_NORESERVE. However, it does guarantee the memory region to be available. The next mmap allocates a portion of the allocated range. There is no MAP_NORESERVE which means this allocation is catered for for swapping in the case of memory shortage. This mapping does use a specific address, so Oracle can use pointers to refer to the contents, because it is certain of the memory address. Also, the MAP_FIXED flag has a side effect, which is used here: any memory mapping done to the address range is silently unmapped from the first (“throw away”) mapping.

Let’s look a bit deeper into the wait event information. For this I changed the probe for function kskthewt in the systemtap script in the following way:

probe process("/u01/app/oracle/product/").function("kskthewt") {
	if ( pid() == target() && register("rsi") == wait_event_nr ) {
		ksuse = register("r13")-4672
		ksuseopc = user_uint16(ksuse + 2098)
		ksusep1 = user_uint64(ksuse + 2104)
		ksusep2 = user_uint64(ksuse + 2112)
		ksusep3 = user_uint64(ksuse + 2120)
		ksusetim = user_uint32(ksuse + 2128)
		printf("kskthewt - wait event#: %u, wait_time:%u, p1:%lu, p2:%lu, p3:%lu\n", ksuseopc, ksusetim, ksusep1, ksusep2, ksusep3)

When running a ‘select count(*) from t2’ again on a freshly started database with a new process with the changed mmap.stp script, this is how the output looks like:

kskthbwt - 524
 mmap, addr 0, size 2097152, protection 3, flags 16418, fd -1, offset 753 return value: 7f1562330000
 mmap, addr 7f1562330000, size 1114112, protection 3, flags 50, fd -1, offset 0 return value: 7f1562330000
kskthewt - wait event#: 524, wait_time:30, p1:2097152, p2:1114112, p3:0
kskthbwt - 524
 mmap, addr 0, size 2097152, protection 3, flags 16418, fd -1, offset 1265 return value: 7f1562130000
 mmap, addr 7f1562130000, size 1114112, protection 3, flags 50, fd -1, offset 0 return value: 7f1562130000
kskthewt - wait event#: 524, wait_time:28, p1:2097152, p2:1114112, p3:0

This looks like the size of memory allocated with the first mmap call for the PGA memory reservation is put in p1, and the size of the allocation of the second “real” memory allocation is put in p2 of the ‘PGA memory operation’ event. One thing that does look weird, is the memory is not unmapped/deallocated (this is a full execution of a SQL, allocated buffers must be deallocated?

Let’s look what happens when I execute the same SQL again:

kskthbwt - 524
 munmap, addr 7f1562130000, size 2097152
 mmap, addr 7f15623b0000, size 589824, protection 0, flags 16434, fd -1, offset 881 return value: 7f15623b0000
kskthewt - wait event#: 524, wait_time:253, p1:0, p2:0, p3:0
kskthbwt - 524
 mmap, addr 7f15623b0000, size 589824, protection 3, flags 50, fd -1, offset 0 return value: 7f15623b0000
kskthewt - wait event#: 524, wait_time:35, p1:589824, p2:0, p3:0
kskthbwt - 524
 mmap, addr 0, size 2097152, protection 3, flags 16418, fd -1, offset 1265 return value: 7f1562130000
 mmap, addr 7f1562130000, size 1114112, protection 3, flags 50, fd -1, offset 0 return value: 7f1562130000
kskthewt - wait event#: 524, wait_time:30, p1:2097152, p2:0, p3:0

Ah! It looks like some memory housekeeping is not done during the previous execution, but is left for the next execution, the execution starts with munmap(), followed by a mmap() call. The first munmap() call deallocates 2 megabyte memory chunk. The next mmap() call is different from the other mmap() calls we have seen so far; we have seen a “throw away”/reservation mmap() call with the memory address set to 0 to let the operating system pick an address for the requested memory chunk, and a mmap() call to truly allocate the reserved memory for usage, which had a memory address set. The mmap() call following munmap() has a memory address set. However, protection is set to 0; this means PROT_NONE, which means the mapped memory can not be read and written. Also the flags number is different, flags 16434 translates to MAP_PRIVATE|MAP_FIXED|MAP_ANONYMOUS|MAP_NORESERVE. As part of releasing PGA memory, it seems some memory is reserved. The wait event parameters are all zero. When p1, p2 and p3 are all zero, it seems to indicate munmap() is called. As we just have seen, memory could be reserved. Also, when p1/2/3 are all zero there is no way to tell how much memory is freed, nor which memory allocation.

The next wait is the timing of a single mmap() call. Actually, the mmap() call allocates the previous mmaped memory, but now with protection set to 3 (PROT_READ|PROT_WRITE), which means the memory is actually usable. The p1 value is the amount of memory mmaped.

The last wait is a familiar one, it is the mmap() call with memory address set to zero, as reservation, and another mmap() call to allocate memory inside the previous “reserved” memory. However, the p1/2/4 values are now NOT set in the same way as we saw earlier: only p1 is non zero, indicating the size of the first mmap() call. Previously, p1 and p2 were set to the sizes of both mmap() calls.

With Oracle version 12.2 there is a new wait event ‘PGA memory operation’. This event indicates memory is allocated or de-allocated. Until now I only saw the system calls mmap() and munmap() inside the ‘PGA memory operation’.

In my previous post, I introduced Intel Pin. If you are new to pin, please follow this link to my previous post on how to set it up and how to run it.

One of the things you can do with Pin, is profile memory access. Profiling memory access using the pin tool ‘pinatrace’ is done in the following way:

$ cd ~/pin/pin-3.0-76991-gcc-linux
$ ./pin -pid 12284 -t source/tools/SimpleExamples/obj-intel64/

The pid is a pid of an oracle database foreground process. Now execute something in the session you attached pin to and you find the ‘pinatrace’ output in $ORACLE_HOME/dbs:

$ ls -l $ORACLE_HOME/dbs
total 94064
-rw-rw----. 1 oracle oinstall     1544 Nov 16 09:40 hc_testdb.dat
-rw-r--r--. 1 oracle oinstall     2992 Feb  3  2012 init.ora
-rw-r-----. 1 oracle oinstall       57 Nov  5 09:42 inittestdb.ora
-rw-r-----. 1 oracle oinstall       24 Nov  5 09:32 lkTESTDB
-rw-r-----. 1 oracle oinstall     7680 Nov  5 09:41 orapwtestdb
-rw-r--r--  1 oracle oinstall 10552584 Nov 17 06:36 pinatrace.out

Please mind memory access generates A LOT of information! The above 11MB is what a ‘select * from dual’ generates (!)

This is how the file looks like:

$ head pinatrace.out
# Memory Access Trace Generated By Pin
0x00007f85c63fe218: R 0x00007fff6fd2c4c8  8          0xcefb615
0x000000000cefb61e: W 0x00007fff6fd2c4f8  8              0x12c
0x000000000cefb621: R 0x00007fff6fd2c4d0  8     0x7f85c5bebd96
0x000000000cefb625: R 0x00007fff6fd2c4d8  8     0x7f85c5bebd96
0x000000000cefb62c: R 0x00007fff6fd2c4e0  8     0x7fff6fd2c570
0x000000000cefb62d: R 0x00007fff6fd2c4e8  8          0xcefb54e

The first field is the function location, the second field is R or W (reading or writing obviously), the third field is the memory location read or written the fourth field is the amount of bits read and the fifth field is prefetched memory.

The function that is used can be looked up using the addr2line linux utility:

$ addr2line -p -f -e /u01/app/oracle/product/ 0x000000000cefb61e
sntpread at ??:?

I looked up the second address from the pinatrace.out file above, and that address belongs to the function sntpread. There is no additional information available for this function (‘at ??:?’). If the address is not available in the oracle executable, a ‘??’ is displayed:

$ addr2line -p -f -e /u01/app/oracle/product/ 0x00007f85c63fe218
?? ??:0

The pinatrace.out file is usable if you know the exact instruction pointer address or the memory location. However, that usage is fairly limited. An example of that is Mahmoud Hatem’s blog on tracing access to a memory location. Wouldn’t it be nice if we can change the functions addresses to function names, and the memory addresses to named memory locations whenever possible?

That’s where I created the pinatrace annotate oracle tool for. This is a little scriptset that contains scripts to generate memory information from the instance, after which the instruction pointer addresses and memory locations of a pinatrace.out file generated by pinatrace are translated to function names and memory area names. Let’s have a look what that means. This is a snippet of a pinatrace.out file:

0x000000000c967e46: R 0x0000000095f69910  8         0x95fcf6b0
0x000000000c967e4d: W 0x00007fff6fd2b2b8  8          0xc967e52
0x000000000c937b32: W 0x00007fff6fd2b2b0  8     0x7fff6fd2bdb0
0x000000000c937b3a: W 0x00007fff6fd2b278  8                0xe
0x000000000c937b41: W 0x00007fff6fd2b298  8         0x95f68ea8
0x000000000c937b45: W 0x00007fff6fd2b270  8                0x1
0x000000000c937b49: W 0x00007fff6fd2b280  8     0x7f85ca1db280
0x000000000c937b4d: R 0x0000000095fcf6bc  2               0x12
0x000000000c937b52: W 0x00007fff6fd2b288  8              0x2c4
0x000000000c937b59: W 0x00007fff6fd2b290  8          0xd8f898c
0x000000000c937b60: W 0x00007fff6fd2b2a0  4               0x73
0x000000000c937b6b: W 0x00007fff6fd2b2a8  4                0x1
0x000000000c937b6e: R 0x00007f85ca1db280  8     0x7f85ca1db280
0x000000000c937b77: R 0x000000000d0a40e4  4                0x1
0x000000000c937b84: R 0x00007f85ca1d43c8  8         0x95dc0e20
0x000000000c937b92: R 0x0000000095dc10b0  8                  0
0x000000000c937ba2: R 0x0000000095fcf6c0  4                0x1
0x000000000c937ba9: R 0x0000000095dc10e0  4                  0
0x000000000c937baf: R 0x000000000cfbe644  4            0x1cffe
0x000000000c937bbc: W 0x0000000095dc10b0  8         0x95fcf6b0
0x000000000c937bc5: R 0x0000000095fcf6b0  8                  0
0x000000000c937bc5: W 0x0000000095fcf6b0  8                0x1
0x000000000c937bca: W 0x00007fff6fd2b260  8                  0
0x000000000c937be1: R 0x00007f85ca1d4290  8     0x7f85ca1a9ca0
0x000000000c937bec: R 0x00007f85ca1ab1c0  4                0x3
0x000000000c937bf3: W 0x0000000095dc0faa  2                0x3
0x000000000c937bf9: R 0x00007f85ca1d43e0  8         0x95f68ea8
0x000000000c937c09: R 0x0000000095f69470  2                  0
0x000000000c937c16: W 0x0000000095dc0fac  2                  0
0x000000000c937c1e: R 0x0000000095dc10e0  4                  0
0x000000000c937c1e: W 0x0000000095dc10e0  4                0x2
0x000000000c937c24: W 0x0000000095dc0fa0  8         0x95fcf6b0
0x000000000c937c28: W 0x0000000095dc0fa8  2                0x8
0x000000000c937c2e: R 0x000000006000a9d8  4                0x1
0x000000000c937c3b: R 0x00007fff6fd2b298  8         0x95f68ea8
0x000000000c937c3f: R 0x00007fff6fd2b2a0  4               0x73
0x000000000c937c42: W 0x0000000095fcf6c8  8         0x95f68ea8
0x000000000c937c46: W 0x0000000095fcf6c4  4               0x73
0x000000000c937c4a: R 0x00007fff6fd2b2a8  4                0x1
0x000000000c937c50: R 0x0000000095fcf6b8  4              0x83e
0x000000000c937c50: W 0x0000000095fcf6b8  4              0x83f
0x000000000c937c5a: W 0x0000000095dc10b0  8                  0
0x000000000c937c65: R 0x00007f85ca1d71d6  1                  0
0x000000000c937c76: R 0x00007fff6fd2b270  8                0x1
0x000000000c937c7a: R 0x00007fff6fd2b290  8          0xd8f898c
0x000000000c937c7e: R 0x00007fff6fd2b288  8              0x2c4
0x000000000c937c82: R 0x00007fff6fd2b280  8     0x7f85ca1db280
0x000000000c937c86: R 0x00007fff6fd2b278  8                0xe
0x000000000c937c8d: R 0x00007fff6fd2b2b0  8     0x7fff6fd2bdb0
0x000000000c937c8e: R 0x00007fff6fd2b2b8  8          0xc967e52

The usefulness of this is limited in this form. The only thing I could derive is that big numbers in the memory access column (‘0x00007fff6fd2ac60’) are probably PGA related, and the numbers between roughly 0x000000006000000 and 0x0000000095dc0fd0 are probably SGA related. After running the annotate tool, it looks like this:

ksl_get_shared_latch:R:0x0000000095fcf6bc(shared pool|permanent memor,duration 1,cls perm shared pool|(child)latch:session idle bit):2
ksl_get_shared_latch:R:0x00007f85ca1db280(pga|Other, pga heap, permanent memory pga|Other, top call heap, free memory):8
ksl_get_shared_latch:R:0x00007f85ca1d43c8(pga|Other, pga heap, permanent memory pga|Other, top call heap, free memory):8
ksl_get_shared_latch:R:0x0000000095dc10b0(shared pool|permanent memor,duration 1,cls perm shared pool|X$KSUPR.KSLLALAQ):8
ksl_get_shared_latch:R:0x0000000095fcf6c0(shared pool|permanent memor,duration 1,cls perm shared pool|(child)latch:session idle bit):4
ksl_get_shared_latch:R:0x0000000095dc10e0(shared pool|permanent memor,duration 1,cls perm shared pool|X$KSUPR.KSLLALOW):4
ksl_get_shared_latch:W:0x0000000095dc10b0(shared pool|permanent memor,duration 1,cls perm shared pool|X$KSUPR.KSLLALAQ):8
ksl_get_shared_latch:R:0x0000000095fcf6b0(shared pool|permanent memor,duration 1,cls perm shared pool|(child)latch:session idle bit):8
ksl_get_shared_latch:W:0x0000000095fcf6b0(shared pool|permanent memor,duration 1,cls perm shared pool|(child)latch:session idle bit):8
ksl_get_shared_latch:R:0x00007f85ca1d4290(pga|Other, pga heap, permanent memory pga|Other, top call heap, free memory):8
ksl_get_shared_latch:R:0x00007f85ca1ab1c0(pga|Other, pga heap, kgh stack pga|Other, pga heap, free memory pga|Other, pga heap, permanent memory):4
ksl_get_shared_latch:W:0x0000000095dc0faa(shared pool|permanent memor,duration 1,cls perm):2
ksl_get_shared_latch:R:0x00007f85ca1d43e0(pga|Other, pga heap, permanent memory pga|Other, top call heap, free memory):8
ksl_get_shared_latch:R:0x0000000095f69470(shared pool|permanent memor,duration 1,cls perm):2
ksl_get_shared_latch:W:0x0000000095dc0fac(shared pool|permanent memor,duration 1,cls perm):2
ksl_get_shared_latch:R:0x0000000095dc10e0(shared pool|permanent memor,duration 1,cls perm shared pool|X$KSUPR.KSLLALOW):4
ksl_get_shared_latch:W:0x0000000095dc10e0(shared pool|permanent memor,duration 1,cls perm shared pool|X$KSUPR.KSLLALOW):4
ksl_get_shared_latch:W:0x0000000095dc0fa0(shared pool|permanent memor,duration 1,cls perm):8
ksl_get_shared_latch:W:0x0000000095dc0fa8(shared pool|permanent memor,duration 1,cls perm):2
ksl_get_shared_latch:R:0x000000006000a9d8(fixed sga|var:kslf_stats_):4
ksl_get_shared_latch:W:0x0000000095fcf6c8(shared pool|permanent memor,duration 1,cls perm shared pool|(child)latch:session idle bit):8
ksl_get_shared_latch:W:0x0000000095fcf6c4(shared pool|permanent memor,duration 1,cls perm shared pool|(child)latch:session idle bit):4
ksl_get_shared_latch:R:0x0000000095fcf6b8(shared pool|permanent memor,duration 1,cls perm shared pool|(child)latch:session idle bit):4
ksl_get_shared_latch:W:0x0000000095fcf6b8(shared pool|permanent memor,duration 1,cls perm shared pool|(child)latch:session idle bit):4
ksl_get_shared_latch:W:0x0000000095dc10b0(shared pool|permanent memor,duration 1,cls perm shared pool|X$KSUPR.KSLLALAQ):8
ksl_get_shared_latch:R:0x00007f85ca1d71d6(pga|Other, pga heap, permanent memory pga|Other, top call heap, free memory):1

So, now you can see the reason I picked a seemingly arbitrary range of lines actually was because that range is the memory accesses of the ksl_get_shared_latch function. This annotated version show a shared latch get for the ‘session idle bit’ latch. It’s also visible the function uses PGA memory, some of it annotated, some of it not, and that most of the shared pool access is for the latch (a latch essentially is a memory range with the function of serialising access to a resource), which is in the shared pool because it’s a child latch. It’s also visible memory belonging to X$KSUPR is read and written (X$KSUPR is the table responsible for V$PROCESS, the fields KSLLALAQ and KSLLALOW are not externalised in V$PROCESS).

Why are a lot of the assumed PGA addresses (the ones like 0x00007fff6fd2b2b8) not annotated? Well, PGA memory allocations are very transient of nature. Because a PGA memory snapshot is made at a certain point in time, this snapshot represents the memory layout of that moment, which has a high probability of having memory deallocated and freed to the operating system. A lot of the SGA/shared pool allocations on the other hand have the intention of re-usability, and thus are not freed immediately after usage, which gives the SGA memory snapshot a good chance of capturing a lot of the memory allocations.

Get the pinatrace oracle annotate tool via github: git clone

Please mind this tool uses the bash shell, it might not work in other shells like ksh.

How to use the tool?
– Use pin with the tool, as described above. Move the the pinatrace.out file from $ORACLE_HOME/dbs to the directory with the script.
Immediately after the trace has been generated (!), execute the following scripts using sqlplus as SYSDBA:
– 0_get_pga_detail.sql (this lists the sessions in the database and requires you to specify the oracle PID of the session)
– 1_generate_memory_ranges.sql
– 2_generate_memory_ranges_xtables.sql
– 3_generate_memory_ranges_pga.sql
This results in the following files: memory_ranges.csv, memory_ranges_pga.csv and memory_ranges_xtables.csv.
Now execute the annotate script:
– ./ pinatrace.out
The script outputs to STDOUT, so if you want to save the annotation, redirect it to a file (> file.txt) or if you want to look and redirect to a file: | tee file.txt.

I hope this tool is useful for your research. If you know a memory area described in the data dictionary that is not included, please drop me a message with the script, then I’ll include it.

This blogpost is an introduction to Intel’s Pin dynamic instrumentation framework. Pin and the pintools were brought to my attention by Mahmoud Hatem in his blogpost Tracing Memory access of an oracle process: Intel PinTools. The Pin framework provides an API that abstracts instruction-set specifics (on the CPU layer). Because this is a dynamic binary instrumentation tool, it requires no recompiling of source code. This means we can use it with programs like the Oracle database executable.
The Pin framework download comes with a set of pre-created tools called ‘Pintools’. Some of these tools are really useful for Oracle investigation and research.

Pin works in a very sophisticated way. The description in the Pin manualis to think of Pin as a JIT (just in time) compiler, where the compiler does not take byte code (as JIT compilation does with Java), but the executable of the process pin is executed against. This means pin inserts itself into the process’ execution. This can be seen when looking at the memory map of such a process:

$ cat /proc/29595/maps | grep -e pin-3.0 -e oracle
00400000-1098a000 r-xp 00000000 fb:02 68469986                           /u01/app/oracle/product/
10b8a000-10bac000 r--p 1058a000 fb:02 68469986                           /u01/app/oracle/product/
10bac000-10e05000 rw-p 105ac000 fb:02 68469986                           /u01/app/oracle/product/
7fa01e04c000-7fa01e31d000 r-xp 00000000 fb:04 67152845                   /home/oracle/pin/pin-3.0-76991-gcc-linux/intel64/lib-ext/
7fa01e51c000-7fa01e530000 r--p 002d0000 fb:04 67152845                   /home/oracle/pin/pin-3.0-76991-gcc-linux/intel64/lib-ext/
7fa01e530000-7fa01e531000 rw-p 002e4000 fb:04 67152845                   /home/oracle/pin/pin-3.0-76991-gcc-linux/intel64/lib-ext/
7fa01e533000-7fa01e71a000 r-xp 00000000 fb:04 954267                     /home/oracle/pin/pin-3.0-76991-gcc-linux/source/tools/SimpleExamples/obj-intel64/
7fa01e91a000-7fa01e91c000 r--p 001e7000 fb:04 954267                     /home/oracle/pin/pin-3.0-76991-gcc-linux/source/tools/SimpleExamples/obj-intel64/
7fa01e91c000-7fa01e91e000 rw-p 001e9000 fb:04 954267                     /home/oracle/pin/pin-3.0-76991-gcc-linux/source/tools/SimpleExamples/obj-intel64/
7fa01e946000-7fa01e9c9000 r-xp 00000000 fb:04 136702                     /home/oracle/pin/pin-3.0-76991-gcc-linux/intel64/runtime/pincrt/
7fa01e9ca000-7fa01e9cc000 r--p 00083000 fb:04 136702                     /home/oracle/pin/pin-3.0-76991-gcc-linux/intel64/runtime/pincrt/
7fa01e9cc000-7fa01e9ce000 rw-p 00085000 fb:04 136702                     /home/oracle/pin/pin-3.0-76991-gcc-linux/intel64/runtime/pincrt/
7fa01e9d6000-7fa01ea04000 r-xp 00000000 fb:04 136694                     /home/oracle/pin/pin-3.0-76991-gcc-linux/intel64/runtime/pincrt/
7fa01ea04000-7fa01ea05000 r--p 0002d000 fb:04 136694                     /home/oracle/pin/pin-3.0-76991-gcc-linux/intel64/runtime/pincrt/
7fa01ea05000-7fa01ea06000 rw-p 0002e000 fb:04 136694                     /home/oracle/pin/pin-3.0-76991-gcc-linux/intel64/runtime/pincrt/
7fa01ea06000-7fa01eac4000 r-xp 00000000 fb:04 136696                     /home/oracle/pin/pin-3.0-76991-gcc-linux/intel64/runtime/pincrt/
7fa01eac5000-7fa01eac8000 r--p 000be000 fb:04 136696                     /home/oracle/pin/pin-3.0-76991-gcc-linux/intel64/runtime/pincrt/
7fa01eac8000-7fa01eaca000 rw-p 000c1000 fb:04 136696                     /home/oracle/pin/pin-3.0-76991-gcc-linux/intel64/runtime/pincrt/
7fa01eacb000-7fa01edb7000 r-xp 00000000 fb:04 100663633                  /home/oracle/pin/pin-3.0-76991-gcc-linux/extras/xed-intel64/lib/
7fa01edb8000-7fa01ee24000 r--p 002ec000 fb:04 100663633                  /home/oracle/pin/pin-3.0-76991-gcc-linux/extras/xed-intel64/lib/
7fa01ee24000-7fa01ee25000 rw-p 00358000 fb:04 100663633                  /home/oracle/pin/pin-3.0-76991-gcc-linux/extras/xed-intel64/lib/
7fa01ee80000-7fa01f385000 r-xp 00000000 fb:04 136689                     /home/oracle/pin/pin-3.0-76991-gcc-linux/intel64/bin/pinbin
7fa01f385000-7fa01f38d000 r--p 00504000 fb:04 136689                     /home/oracle/pin/pin-3.0-76991-gcc-linux/intel64/bin/pinbin
7fa01f38d000-7fa01f394000 rw-p 0050c000 fb:04 136689                     /home/oracle/pin/pin-3.0-76991-gcc-linux/intel64/bin/pinbin
7fa01f54e000-7fa01f6d4000 r-xp 00000000 fb:02 212053053                  /u01/app/oracle/product/
7fa01f6d4000-7fa01f8d3000 ---p 00186000 fb:02 212053053                  /u01/app/oracle/product/

Here you see this process an oracle foreground (rows 1-3), after which we see a lot of stuff that pin pushed into the process’ address space.

For this reason, it’s vitally important to use and run pin as the same user as the process you want to run pin against. The way pin works is that, upon execution of pin, the pin executable inserts itself into the process’ address space, gains control and then tries to load necessary libraries. If it can’t find these libraries, it will send a SIGKILL to the process, effectively stopping it!

In order to obtain pin, go to the pin homepage, downloads, linux and select the ‘gcc compiler kit’. This is a zipped tarball. Upload that to a linux server as the database owner, usually ‘oracle’. I created a directory ‘pin’ in which I put the tarball, and extracted it (tar xzf). The next step is to compile the tools that come with pin:

$ cd pin-3.0-76991-gcc-linux/source/tools
$ make

This will output a lot of stuff on your screen and compiles everything in the tools directory.

Now that we have seen an introduction, let’s use a Pin tool to do something useful! One such useful is the tool ‘DebugTrace’. First obtain the process id (pid) from an Oracle server foreground process. Then execute the pin tool against this process:

$ pwd
$ ./pin -pid 2407 -t source/tools/DebugTrace/obj-intel64/

Upon execution, nothing is returned, because the majority of the things are happening in the process pin is run against (2407 in this case). If no parameters are specified, the pintool will create a file in the current working directory of the process it is run against. For an oracle foreground process, the current working directory is $ORACLE_HOME/dbs:

$ ls -l /proc/2407/cwd
lrwxrwxrwx 1 oracle oinstall 0 Nov 17 01:45 /proc/2407/cwd -> /u01/app/oracle/product/

For the sake of testing this, I ran ‘select * from dual’ in the oracle database foreground process.

Let’s look what this produces! This is how my dbs directory looks like:

$ ls -ltr
total 2332
-rw-r--r--. 1 oracle oinstall    2992 Feb  3  2012 init.ora
-rw-r-----. 1 oracle oinstall      24 Nov  5 09:32 lkTESTDB
-rw-r-----. 1 oracle oinstall    7680 Nov  5 09:41 orapwtestdb
-rw-r-----. 1 oracle oinstall      57 Nov  5 09:42 inittestdb.ora
-rw-rw----. 1 oracle oinstall    1544 Nov 16 09:40 hc_testdb.dat
-rw-r--r--  1 oracle oinstall 2361460 Nov 17 01:49 debugtrace.out

You can see the debugtrace.out file is what is produced by the pin tool, and it look rather big (remember I only ran ‘select * from dual’)!

Let’s look into the file:

$ less debugtrace.out
         3 @@@ return underflow
Return 0x00007f706b664218 /lib64/ returns: 0x12c
        14 @@@ return underflow
Return 0x000000000cefb62d /u01/app/oracle/product/ returns: 0
        31 @@@ return underflow
Return 0x000000000cefb580 /u01/app/oracle/product/ returns: 0
       119 @@@ return underflow
Return 0x000000000cedcc9d /u01/app/oracle/product/ returns: 0
       122 @@@ return underflow
Return 0x000000000cedc90a /u01/app/oracle/product/ returns: 0
       152 Call 0x000000000ceeb746 /u01/app/oracle/product/ -> 0x0000000005b170d0 /u01/app/oracle/product/, 0x7f706ae51da0, ...)
       156 Tailcall 0x0000000005b170de /u01/app/oracle/product/ -> 0x0000000005b170c0 /u01/app/oracle/product/, 0x7f706ae51da0, ...)
       158 | Call 0x0000000005b170c1 /u01/app/oracle/product/ -> 0x0000000005b1dcb0 /u01/app/oracle/product/, 0x7f706ae51da0, ...)
       172 | Return 0x0000000005b20097 /u01/app/oracle/product/ returns: 0x7ffeef95c305
       174 Return 0x0000000005b170c7 /u01/app/oracle/product/ returns: 0x7ffeef95c305
       197 @@@ return underflow
Return 0x000000000ceeb7ad /u01/app/oracle/product/ returns: 0
       213 Call 0x000000000cb5d8b0 /u01/app/oracle/product/ -> 0x000000000c940560 /u01/app/oracle/product/, 0x7f706ae51da1, ...)
       268 | Call 0x000000000c94066e /u01/app/oracle/product/ -> 0x000000000cd585c0 /u01/app/oracle/product/, 0x7f706ae51da1, ...)
       278 | | Call 0x000000000cd5863b /u01/app/oracle/product/ -> 0x000000000b13ef90 /u01/app/oracle/product/, 0x7f706ae51da1, ...)
       286 | | | Call 0x000000000b13efa3 /u01/app/oracle/product/ -> 0x0000000000ba6860 /u01/app/oracle/product/, 0x7ffeef95beb0, ...)
       293 | | | | Call 0x00007f706b188dcb /lib64/ -> 0x00007ffeef9ab8a0 clock_gettime(0x1, 0x7ffeef95beb0, ...)

I think this is rather exciting! This is a FULL function call trace, indented by call depth!
The first lines in the trace are returns, and these returns produce a ‘return underflow’, which is because these functions have been called before the trace was put on the process.
This trace shows:
Calling: the address and function where the function call is made, the address and function that is called and first two arguments of the function.
Returning: the address and function where the return is executed, and the returncode.
Tailcall: this is a function that is called as the last instruction in the current function. This means it returns as part of the current function, which is shown by not indenting this call further (call, tailcall and return are all indented at the same level).

Please mind currently I am not aware of a way to remove the pin framework and tool from a process address space. Also mind that because of how this works (a lot of intercepting in the execution flow), it will slow down the process significantly. Use this tool and the above described techniques at your own risk.

This blogpost is about the Oracle redo log structures and redo efficiency in modern Oracle databases. Actually, a lot of subtle things changed surrounding redo (starting from Oracle 10 actually) which have gone fairly unnoticed. One thing the changes have gone unnoticed for is the Oracle documentation, the description of redo in it is an accurate description for Oracle 9, not how it is working in Oracle 10 or today in Oracle

My test environment is a virtual machine with Oracle Linux 7.2 and Oracle, and a “bare metal” server running Oracle Linux 6.7 and Oracle Versions are important, as things can change between versions.

Multi threaded redo and dynamic strands
One of the prominent changes to the redo mechanism is what is referred to as ‘multi threaded redo’ or ‘multiple log buffers’. The way the redo structures are described most of the time is the redo log buffer being in the SGA as a single buffer to which change vectors are written (in a circular fashion), protected by a latch (redo allocation latch). That is not an accurate description since Oracle 9.2.

Starting from Oracle 9.2, multiple log buffers can be and most probably are created. At least in Oracle 12.1 you get 2 redo log buffers. Each of these buffers is protected by a redo allocation latch. The benefit of multiple log buffers is that copying redo into the log buffer is not strictly serialised. However, this has some consequences too. One of the consequences is redo potentially could be not written in strict SCN order by the log writer or its worker processes.

You can see the log buffers in the view x$kcrfstrand:

SYS@testdb AS SYSDBA> select strand_size_kcrfa from x$kcrfstrand where ptr_kcrf_pvt_strand = hextoraw(0) and pnext_buf_kcrfa_cln != hextoraw(0);


Both values added together exactly matches the size of the log buffer:

SYS@testdb AS SYSDBA> @parms
Enter value for parameter: log_buffer
old  20: where name like nvl('%&parameter%',name)
new  20: where name like nvl('%log_buffer%',name)
Enter value for isset:
old  21: and upper(isset) like upper(nvl('%&isset%',isset))
new  21: and upper(isset) like upper(nvl('%%',isset))
Enter value for show_hidden:
old  22: and flag not in (decode('&show_hidden','Y',3,2))
new  22: and flag not in (decode('','Y',3,2))

NAME                         VALUE                                                                  ISDEFAUL ISMODIFIED ISSET
---------------------------- ---------------------------------------------------------------------- -------- ---------- ----------
log_buffer                   134217728                                                              FALSE    FALSE      TRUE

The parameters that handle dynamic redo strands are “hidden”:

SYS@testdb AS SYSDBA> @parms
Enter value for parameter: log_parallelism
old  20: where name like nvl('%&parameter%',name)
new  20: where name like nvl('%log_parallelism%',name)
Enter value for isset:
old  21: and upper(isset) like upper(nvl('%&isset%',isset))
new  21: and upper(isset) like upper(nvl('%%',isset))
Enter value for show_hidden: Y
old  22: and flag not in (decode('&show_hidden','Y',3,2))
new  22: and flag not in (decode('Y','Y',3,2))

NAME                            VALUE                                                                  ISDEFAUL ISMODIFIED ISSET
------------------------------- ---------------------------------------------------------------------- -------- ---------- ----------
_log_parallelism_dynamic        TRUE                                                                   TRUE     FALSE      FALSE
_log_parallelism_max            2                                                                      TRUE     FALSE      FALSE

The mechanisms can further be investigated using fixed SGA variables:

SYS@testdb AS SYSDBA> oradebug setmypid
Statement processed.
SYS@testdb AS SYSDBA> oradebug dumpvar sga kcrf_max_strands
uword kcrf_max_strands_ [0600283C8, 0600283CC) = 00000002
SYS@testdb AS SYSDBA> oradebug dumpvar sga kcrf_actv_strands
uword kcrf_actv_strands_ [0600283E0, 0600283E4) = 00000001

Note: the output of oradebug is hexadecimal (!). Of course here it doesn’t matter since the values are smaller than 10.
This shows the maximum number of public redo strands is 2 (kcrf_max_strands), as we have seen with x$kcrfstrand, and the number of active strands is 1 (kcrf_actv_strands), which makes sense in my test database, since I am the sole user of the database.
The maximum number of public redo strands seems to be at least 2, some limited research shows the value is set by the formula CPU_COUNT/16 for systems with higher CPU counts. Some sources report that the value for _log_parallelism_max will be reduced to CPU_COUNT if it is set higher manually.

When I count the number of redo allocation latches, I see something odd:

SYS@testdb AS SYSDBA> select count(*) from v$latch_children where name = 'redo allocation';


A redo allocation latch protects a log strand. I just showed my system has actually two public log strands. Why do I have 20 redo allocation latches if my system has a fixed number of public redo buffers? This finding leads to private redo strands.

Private strands
Outside of multi threaded public redo buffers, Oracle introduced another optimisation for redo, which are private strands. Private strands optimise redo generation by letting a session create the change vectors for a transaction directly in a private redo buffer in the SGA, instead of generating the change vectors in the PGA, and later copying these to the public redo strand. Some restrictions apply on the use of it, although I am not aware of a concrete list of restrictions and limitations (the closest thing to such a list would be the table x$ktiff shown further below with in-memory undo, which is directly connected to private strands). Whenever a private strand can not be used, Oracle reverts to the ‘classic’ mechanism of redo generation using the public redo log buffer. Each private strand is also protected by a redo allocation latch, just like the public strands.

The private strands SGA buffers are allocated at startup time:

SYS@testdb AS SYSDBA> select pool, name, bytes from v$sgastat where name like 'private strands';

POOL         NAME                            BYTES
------------ -------------------------- ----------
shared pool  private strands               2451456

It seems the number of private strands allocated from SGA (and as a consequence the number of redo allocation latches on top of the public strand number) is depended on the transactions database parameter, and takes 10% (_log_private_parallelism_mul) of the transactions multiplied by 129KB plus 4KB overhead:

SYS@testdb AS SYSDBA> select trunc(value * KSPPSTVL / 100) * (129+4) * 1024 "private strands"
                      from (select value from v$parameter where name = 'transactions') a,
                           (select val.KSPPSTVL
                            from sys.x$ksppi nam, sys.x$ksppsv val
                            where nam.indx = val.indx AND nam.ksppinm = '_log_private_parallelism_mul') b;

private strands

However, that is not all that is in play for private strands. Once the private strands are allocated, and after every logswitch, the number of usable private strands is determined by the size of the current online redologfile minus the size of the log buffer and the 5% multiplier (_log_private_mul):

SYS@testdb AS SYSDBA> select trunc(((select bytes from v$log where status = 'CURRENT') - (select to_number(value) from v$parameter where name = 'log_buffer'))*
                             (select to_number(val.KSPPSTVL)
                              from sys.x$ksppi nam, sys.x$ksppsv val
                              where nam.indx = val.indx AND nam.ksppinm = '_log_private_mul') / 100 / 66560)
                             as "private strands"
                       from dual;

private strands

Yes, that is a negative number. Let’s look at the number of private strands in this instance:

SYS@testdb AS SYSDBA> select indx,strand_size_kcrfa from x$kcrfstrand where last_buf_kcrfa = '00';

no rows selected

The instance takes the least of the two calculations above for the number of private strands. That explains my test instance not having any private strands! The size of my log buffer is 128M, the size of my online redologfiles is 100M. This shows that sizing the online redologfiles too small can lead to inefficiencies (!!).

In-Memory Undo
Another mechanism that has been changed in order to optimise generating redo is in-memory undo. Actually in-memory undo is directly tied to private redo strands. The idea of in-memory undo is to store the undo (read consistent related) part of a transaction in an in-memory undo buffer until it is committed or rolled back. The in-memory undo pool can be seen in the shared pool using:

SYS@testdb AS SYSDBA> select pool, name, bytes from v$sgastat where name = 'KTI-UNDO';

POOL         NAME                                                    BYTES
------------ -------------------------------------------------- ----------
shared pool  KTI-UNDO                                              1274976

The in-memory undo pool is externalised through x$ktifp. This view shows the in-memory buffers and their sizes:

SYS@testdb AS SYSDBA> select ktifpno, ktifppsi from x$ktifp;

---------- ----------
         0      65535
        17      65535

The number of in-memory undo buffers is determined by the transactions parameter:

SYS@testdb AS SYSDBA> show parameter transactions

NAME                                 TYPE        VALUE
------------------------------------ ----------- ------------------------------
transactions                         integer     189
transactions_per_rollback_segment    integer     5

And defaults to 10% of the transactions. Every in-memory undo buffer is protected by an in-memory undo latch:

SYS@testdb AS SYSDBA> select count(*) from v$latch_children where name like 'In memory undo latch';

In memory undo can be switched on and off on at instance-wide level or at the session level using the ‘_in_memory_undo’ parameter, and is true/on by default. Switching off un-memory undo switches off private strands too.

In fact some of these optimisations are announced in the alert log file during startup of an instance:

ILAT =18

BR means batched redo mode, which means the private strands feature is turned on (I found that with the too small online redologfiles described above, which caused on private private strands to be made available, the database still announces ‘BR’).
ILAT means number of in-memory pools and latches.

Because in-memory undo is directly tied to private redo strands, the usage of private strands and in-memory undo can be monitored using the view x$ktiff:

SYS@testdb AS SYSDBA> select ktiffcat, ktiffflc from x$ktiff;

KTIFFCAT															   KTIFFFLC
-------------------------------------------------------------------------------------------------------------------------------- ----------
Undo pool overflow flushes														  0
Stack cv flushes															  1
Multi-block undo flushes														  0
Max. chgs flushes															  0
NTP flushes																  0
Contention flushes															  0
Redo pool overflow flushes														  0
Logfile space flushes															  0
Multiple persistent buffer flushes													  0
Bind time flushes															  0
Rollback flushes															  2
Commit flushes															       1694
Recursive txn flushes															  0
Redo only CR flushes															  0
Ditributed txn flushes															  0
Set txn use rbs flushes 														  0
Bitmap state change flushes														  0
Presumed commit violation														  0
Securefile direct-write lob update flushes												  0
Unknown 																  0
Unknown 																  0

This view shows the reasons for flushing the in-memory undo buffers.

Throw-away undo
The reason that the database is mentioning ‘IMODE’ is there is actually another mode one besides ‘BR’, ‘TUA’. The ‘TUA’ abbreviation is a bit weird (Throw Undo Away vs. Throw-away undo?), but it means the database is set to ‘throw-away undo mode’. This is accomplished by setting the undo_retention parameter to a value of 3 or lower, and setting the undocumented parameter _imu_pools to a value greater than 3. Once this is done, and the instance is restarted, it will be reported in the alert.log file as follows:

ILAT =189

The purpose of throw-away undo mode is to throw away undo instead of applying (the name already hints that :-)). By enabling throw-away undo, some of the actions normally executed are not done anymore. Such actions: are applying undo (not all undo is skipped), which reduces block changes, read, writes, buffer activity and reduces redo activity. By enabling this mode, and thus not performing some of the actions that are normally done, there are some consequences: it will significantly increase the risk of ORA-1555 snapshot too old errors. Also a lot of features of the database are not compatible with ‘TUA mode’: flashback database, cluster_database=TRUE, supplemental logging to name a few. However, it looks to me as a mode designed for a specific purpose: being able to do massive loads or changes at a faster rate, and with some consequences. I would add that this mode should always be a temporal setting, and once the desired action is completed, the ‘TUA’ mode should be reverted back to normal behaviour by resetting _imu_pools and increasing undo_retention back to your business transaction needs.

The aim for this blogpost is to describe private strands, in-memory undo and throw-away undo. The name of the blogpost is actually a hat-tip to mainly the paper from Stephan Haisley, which has most of what is written here covered. So in that sense, it is a ‘redo’ of his paper, although I tested this all on Oracle version database instance.

Private strands and in-memory undo reduce the work needed when changes are applied to the database. You might want to make sure you are using these features, especially since too small online redologfiles can prevent private strands from being made available. I come across a lot of databases which have inappropriately sized online redologfiles.

Throw-away undo mode, or TUA mode, is a mode which can reduce the time spend on bulk actions. However, anything comes at a price. For TUA mode it means the database looses some of its durability properties. However, if this is used for importing huge amounts of data, or doing mass changes without any session concurrency, this could be a welcome time saver.

Stephan Haisley’s 2008 presentation ‘Redo and Undo Optimisations in Oracle 10g’
Tanel Põder’s presentation ‘Performance and Scalability Improvements in Oracle 10g and 11g’
Jonathan Lewis’ write-ups about private strands and ‘Oracle Core essential internals for DBAs and Developers’.
Twitter: @piontekdd, @martinberx, @westendwookie
Thanks: Mauro Pagano

There are many posts about the amount of memory that is taken by the Oracle database executables and the database SGA and PGA. The reason for adding yet another one on this topic is a question I recently gotten, and the complexities which surrounds memory usage on modern systems. The intention for this blogpost is to show a tiny bit about page sharing of linux for private pages, then move on to shared pages, and discuss how page allocation looks like with Oracle ASMM (sga_target or manual memory).

The version of linux in this blogpost is Oracle Linux 7.2, using kernel: 4.1.12-37.6.3.el7uek.x86_64 (UEK4)
The version of the Oracle database software is (july 2016).

Memory usage of virtual memory systems is complicated. For that reason I see a lot of people getting very confused about this topic. Let me state a very simple rule: the memory actively being used on a system should fit in physical memory. Swap (a file or partition), increases total virtual memory, but really only is a safety net for saving your system from an out of memory situation at the cost of moving pages from and to disk. Because modern linux kernels have swappiness (willingness to swap) to a non-zero value, it’s not uncommon to have some swap being used, despite physical memory not being oversubscribed. A system stops performing as soon as paging in and out starts to occur, and for that reason should not happen.

1. Private pages for linux executables
When an executable is executed on linux from the shell, the shell executes a fork() call to create a new process, which is implemented as a clone() system call on linux. Using the clone() system call, the virtual memory space of the newly created process is shared (readonly) with it’s parent. This includes the private allocations! Once the child process needs to write in it’s memory space, it will page fault and create it’s own version, abandoning the version of its parent.

Can we actually prove this is happening? Yes, the /proc/ filesystem gives an insight to a process’ virtual memory space.
Let’s start off with a very simple example: we execute ‘cat /proc/self/maps’ to see our own address space:

[oracle@oracle-linux ~]$ cat /proc/self/maps
00400000-0040b000 r-xp 00000000 fb:00 201666243                          /usr/bin/cat
0060b000-0060c000 r--p 0000b000 fb:00 201666243                          /usr/bin/cat
0060c000-0060d000 rw-p 0000c000 fb:00 201666243                          /usr/bin/cat
00e41000-00e62000 rw-p 00000000 00:00 0                                  [heap]
7f69729be000-7f6978ee5000 r--p 00000000 fb:00 576065                     /usr/lib/locale/locale-archive
7f6978ee5000-7f6979099000 r-xp 00000000 fb:00 522359                     /usr/lib64/
7f6979099000-7f6979298000 ---p 001b4000 fb:00 522359                     /usr/lib64/
7f6979298000-7f697929c000 r--p 001b3000 fb:00 522359                     /usr/lib64/
7f697929c000-7f697929e000 rw-p 001b7000 fb:00 522359                     /usr/lib64/
7f697929e000-7f69792a3000 rw-p 00000000 00:00 0
7f69792a3000-7f69792c4000 r-xp 00000000 fb:00 522352                     /usr/lib64/
7f69794b9000-7f69794bc000 rw-p 00000000 00:00 0
7f69794c3000-7f69794c4000 rw-p 00000000 00:00 0
7f69794c4000-7f69794c5000 r--p 00021000 fb:00 522352                     /usr/lib64/
7f69794c5000-7f69794c6000 rw-p 00022000 fb:00 522352                     /usr/lib64/
7f69794c6000-7f69794c7000 rw-p 00000000 00:00 0
7ffdab1c7000-7ffdab1e8000 rw-p 00000000 00:00 0                          [stack]
7ffdab1ea000-7ffdab1ec000 r--p 00000000 00:00 0                          [vvar]
7ffdab1ec000-7ffdab1ee000 r-xp 00000000 00:00 0                          [vdso]
ffffffffff600000-ffffffffff601000 r-xp 00000000 00:00 0                  [vsyscall]

Here’s a lot to see, but we see the cat executable at 0x00400000. The reason for three memory allocations are (linux/ELF) executables uses different sections with specific functions. A full overview of these can be obtained using the readelf executable. A simpler overview of an executable, which matches the above three memory allocations for the cat executable can be obtained using ‘size -B’ (the size executable, -B means ‘berkeley style’, which is default):

[oracle@oracle-linux ~]$ size -B /usr/bin/cat
   text	   data	    bss	    dec	    hex	filename
  43905	   1712	   2440	  48057	   bbb9	/usr/bin/cat

This describes the three memory sections an linux executable can have: text (the machine instructions, alias ‘the program’), data (all initialised variables declared in the program) and BSS (uninitialised data).
The first section always is the text allocation (not sure if it’s impossible to have the text section not being the first allocation, I have never seen it different). If you look at the memory flags, ‘r-xp’, this totally makes sense: ‘r-‘ meaning: read(only, followed by a’-‘ instead of a ‘w’), ‘x’: executable and ‘p’: this is a private allocation. The next allocation is the data section. We don’t execute variables, we read them, which is reflected in the flags: ‘r–p’. But what if we change the value of a variable? That is where the third section is for: changed values of initialised variables. This can be seen from the flag of this section: ‘rw-p’, read, write and private. The fourth allocation lists [heap], this is a mandatory allocation in every process’ memory space, which holds (small) memory allocations, this is NOT the BSS section. In this case, the BSS section does not seem to be allocated.

By having memory allocations for /usr/lib64/ we can see this is a dynamically linked executable. You can also see this by executing ‘file’ on the executable:

[oracle@oracle-linux ~]$ file /usr/bin/cat
/usr/bin/cat: ELF 64-bit LSB executable, x86-64, version 1 (SYSV), dynamically linked (uses shared libs), for GNU/Linux 2.6.32, BuildID[sha1]=3207edc47638918ceaeede21947a20a4a496cf63, stripped

If a linux executable is dynamically linked, you can see the libraries that are loaded by the dynamic linker/loader using the ldd utility:

[oracle@oracle-linux ~]$ ldd /usr/bin/cat =>  (0x00007ffceb3e4000) => /lib64/ (0x00007fd46fb7e000)
       	/lib64/ (0x000055d5253c9000)

This output shows the dynamic loader (/lib64/, and two libraries the dynamic loader loads: and The first one, libc, is the standard C library. The second one, linux-vdso is for virtual dynamic shared object, which is an optimisation for certain system calls to be executed in user space (notably gettimeofday()).
The other allocations that exist in our example are anonymous mappings (usually done by programs using the mmap() call):

7f69794c6000-7f69794c7000 rw-p 00000000 00:00 0

And some allocations for system purposes, like stack, var, vdso and vsyscall.

Now that you have become familiar with some basic linux memory address space specifics, let’s take it a little further. It’s possible to see more about the memory segments using the proc filesystem smaps file:

[oracle@oracle-linux ~]$ cat /proc/self/smaps
00400000-0040b000 r-xp 00000000 fb:00 201666243                          /usr/bin/cat
Size:                 44 kB
Rss:                  44 kB
Pss:                  44 kB
Shared_Clean:          0 kB
Shared_Dirty:          0 kB
Private_Clean:        44 kB
Private_Dirty:         0 kB
Referenced:           44 kB
Anonymous:             0 kB
AnonHugePages:         0 kB
Swap:                  0 kB
KernelPageSize:        4 kB
MMUPageSize:           4 kB
Locked:                0 kB
VmFlags: rd ex mr mw me dw sd
0060b000-0060c000 r--p 0000b000 fb:00 201666243                          /usr/bin/cat
Size:                  4 kB
Rss:                   4 kB

Per allocation there are a lot of properties to be seen. ‘Size’ is the full size, ‘Rss’ is the resident set size, alias the amount of data of this segment that is truly resident for this process in it’s address space. ‘Pss’ is fairly unknown, and is the proportional size of this segment. The way it is proportional is that if pages in this allocation are shared with other processes, the size of these pages are divided by the number processes it is shared with. In this case, we have loaded the text segment of the cat executable into the process’ address space, which all is resident (size and rss are the same) and it’s not shared with any process (rss equals pss). There are many more properties, but these are out of scope for this blogpost.

Now let’s move on to Oracle. If you look at the maps output of the pmon process for example, you’ll see:

[oracle@oracle-linux 14153]$ cat maps
00400000-1096e000 r-xp 00000000 fb:03 67209358                           /u01/app/oracle/product/
10b6d000-10b8f000 r--p 1056d000 fb:03 67209358                           /u01/app/oracle/product/
10b8f000-10de8000 rw-p 1058f000 fb:03 67209358                           /u01/app/oracle/product/
10de8000-10e19000 rw-p 00000000 00:00 0
1190f000-11930000 rw-p 00000000 00:00 0                                  [heap]

Here we see the Oracle executable, with a text segment, a readonly data segment and a read/write data segment, and we see an anonymous mapping directly following the data segments. That’s the BSS segment!
However, what is more interesting to see, is the properties of the distinct memory allocations in smaps:

[oracle@oracle-linux 14153]$ cat smaps
00400000-1096e000 r-xp 00000000 fb:03 67209358                           /u01/app/oracle/product/
Size:             267704 kB
Rss:               40584 kB
Pss:                 819 kB
Shared_Clean:      40584 kB
Shared_Dirty:          0 kB
Private_Clean:         0 kB
Private_Dirty:         0 kB
Referenced:        40584 kB
Anonymous:             0 kB
AnonHugePages:         0 kB
Swap:                  0 kB
KernelPageSize:        4 kB
MMUPageSize:           4 kB
Locked:                0 kB
VmFlags: rd ex mr mw me dw sd
10b6d000-10b8f000 r--p 1056d000 fb:03 67209358                           /u01/app/oracle/product/
Size:                136 kB
Rss:                 124 kB

If we look at the text segment for the oracle binary, we see the total text size is 267704 kB (size), but resident (truly available for this process in its address space) is only 40584 kB (rss), and because the oracle executable’s text segment is shared with a lot of processes, the proportional size is only 819 kB (pss).

If you want to understand how much memory is taken in the system, the size is telling the total size of the segment, but it doesn’t say anything on true memory usage. The rss size tells the amount of pages for the segment that is paged in to the address space of every process, and can (and is, for oracle) different for every process. The pss size is the proportional size for every process. Probably the only way to tell the true amount of memory taken by executables and libraries is to add up all the pss sizes. Any other value only tells something about the process’ point of view on memory usage, but not overall, true consumed space because that would lead to counting too much.

This is different for anonymous allocations. Since anonymous allocations are created when a process is run, I’ve only seen them initialised purely private. For that reason rss and pss sizes are equal, because every process initialises it strictly for itself. This too works in a lazy allocation way. When memory is allocated, the size is defined, but is only really allocated once it’s truly used, which is expressed by a difference between size and rss.

2. shared pages
The Oracle databases relies on shared caches and data structures, which are put into what is called the SGA, the system global area. The main components of the SGA are the shared pool (shared structures), log buffer (change vectors to be written to disk to persist changes) and the buffer cache, amongst others. With any memory management option (manual management, ASMM (automatic shared memory management, sga_target) and AMM (automatic memory management, memory_target)) there is a SGA. Depending on the memory option, these are visible in a different way.

When manual memory or ASMM is used, shared memory is allocated as system V shared memory. The ‘classic’ way of looking at system V shared memory is using ipcs -m (m is for shared memory, you can also use s for semaphores and q for message queues):

[oracle@oracle-linux ~]$ ipcs -m

------ Shared Memory Segments --------
key        shmid      owner      perms      bytes      nattch     status
0x00000000 655360     oracle     600        2932736    124
0x00000000 688129     oracle     600        905969664  62
0x00000000 720898     oracle     600        139673600  62
0x5f921964 753667     oracle     600        20480      62

Please mind that if you have more than one instance active, or an ASM instance active, you will see more shared memory allocations.
Apparently, the oracle database allocates a couple of shared memory segments. If you want to understand what these memory allocations are for, you can use the oradebug ipc command to see what their functions are:

SQL> oradebug setmypid
Statement processed.
SQL> oradebug ipc
IPC information written to the trace file

This generates a trace file in the ‘trace’ directory in the diagnostics destination. Here is how this looks like (partial output with content of interest to this blogpost):

 Area #0 `Fixed Size' containing Subareas 2-2
  Total size 00000000002cbe70 Minimum Subarea size 00000000
   Area  Subarea    Shmid    Segment Addr    Stable Addr    Actual Addr
      0        2   655360 0x00000060000000 0x00000060000000 0x00000060000000
               Subarea size     Segment size   Req_Protect  Cur_protect
                          00000000002cc000 00000000002cc000 default       readwrite
 Area #1 `Variable Size' containing Subareas 0-0
  Total size 0000000036000000 Minimum Subarea size 00400000
   Area  Subarea    Shmid    Segment Addr    Stable Addr    Actual Addr
      1        0   688129 0x00000060400000 0x00000060400000 0x00000060400000
               Subarea size     Segment size   Req_Protect  Cur_protect
                          0000000036000000 0000000036000000 default       readwrite
 Area #2 `Redo Buffers' containing Subareas 1-1
  Total size 0000000008534000 Minimum Subarea size 00001000
   Area  Subarea    Shmid    Segment Addr    Stable Addr    Actual Addr
      2        1   720898 0x00000096400000 0x00000096400000 0x00000096400000
               Subarea size     Segment size   Req_Protect  Cur_protect
                          0000000008534000 0000000008534000 default       readwrite
 Area #3 `skgm overhead' containing Subareas 3-3
  Total size 0000000000005000 Minimum Subarea size 00000000
   Area  Subarea    Shmid    Segment Addr    Stable Addr    Actual Addr
      3        3   753667 0x0000009ec00000 0x0000009ec00000 0x0000009ec00000
               Subarea size     Segment size   Req_Protect  Cur_protect
                          0000000000005000 0000000000005000 default       readwrite

The first allocation is ‘fixed size’, alias the fixed SGA, the second allocation is the ‘variable size’, which contains the shared pool and the buffercache, the third allocation is the ‘redo buffers’ and the fourth is the ‘skgm overhead’ alias the index into the shared memory structures for this instance.

Because any memory allocation is visible in maps and smaps, this method can be used for shared memory too, to see how the shared memory segments are mapped into the process address space. All oracle database server processes have the shared memory segments for the instance mapped into their address space. The usage is different per process, so the amount of shared memory paged into the address space will be different:

12bcd000-12bee000 rw-p 00000000 00:00 0                                  [heap]
60000000-60001000 r--s 00000000 00:05 655360                             /SYSV00000000 (deleted)
60001000-602cc000 rw-s 00001000 00:05 655360                             /SYSV00000000 (deleted)
60400000-96400000 rw-s 00000000 00:05 688129                             /SYSV00000000 (deleted)
96400000-9e934000 rw-s 00000000 00:05 720898                             /SYSV00000000 (deleted)
9ec00000-9ec05000 rw-s 00000000 00:05 753667                             /SYSV5f921964 (deleted)
7f473004e000-7f47301d4000 r-xp 00000000 fb:02 212635773                  /u01/app/oracle/product/

Shared memory is easily identified by the ‘s’, at which “normal” private memory mappings have ‘p’. If you want to know more about the process’ perspective of the shared memory, we can use smaps, just like with private memory mappings (virtual memory space of pmon):

60000000-60001000 r--s 00000000 00:05 655360                             /SYSV00000000 (deleted)
Size:                  4 kB
Rss:                   0 kB
Pss:                   0 kB
Shared_Clean:          0 kB
Shared_Dirty:          0 kB
Private_Clean:         0 kB
Private_Dirty:         0 kB
Referenced:            0 kB
Anonymous:             0 kB
AnonHugePages:         0 kB
Swap:                  0 kB
KernelPageSize:        4 kB
MMUPageSize:           4 kB
Locked:                0 kB
VmFlags: rd sh mr mw me ms sd
60001000-602cc000 rw-s 00001000 00:05 655360                             /SYSV00000000 (deleted)
Size:               2860 kB
Rss:                 392 kB
Pss:                  36 kB
Shared_Clean:          0 kB
Shared_Dirty:        372 kB
Private_Clean:         0 kB
Private_Dirty:        20 kB
Referenced:          392 kB
Anonymous:             0 kB
AnonHugePages:         0 kB
Swap:                  0 kB
KernelPageSize:        4 kB
MMUPageSize:           4 kB
Locked:                0 kB
VmFlags: rd wr sh mr mw me ms sd

These two shared memory segments are belonging to the fixed sga. The reason for two segments is the first page (0x1000 equals 4096, alias a single linux page) is readonly (r–s). The other fixed SGA segment is read write (rw-s). Here we see that from the process’ perspective it really doesn’t matter much if a piece of mapped memory is shared or private; it’s exactly handled the same way, which means the full segment is mapped into the process’ virtual memory space, but only once pages are touched (alias truly used), the process registers the address in its pagetable, and the pages become resident (as can be seen in the difference between the total size and the rss). The sole purpose of shared memory is it is shared between process. That the pages are shared is very well visible with the difference between rss and pss size. Its also easy to spot this shared memory segment is created from small pages; MMUPageSize and KernelPageSize is 4kB.

However, this yields an interesting question: shared memory does not belong to any single process. Does that mean that if a shared memory segment is created, it is truly allocated, or can shared memory be lazy allocated as well? Please mind that above statistics are the process’ perspective, not the kernel’s perspective.

One way to see the state of shared memory system wide, is using the ‘-u’ flag with the ipcs command:

[oracle@oracle-linux [testdb] ~]$ ipcs -mu

------ Shared Memory Status --------
segments allocated 4
pages allocated 256005
pages resident  255684
pages swapped   0
Swap performance: 0 attempts   	 0 successes

This is a really useful view! What we can see from the output from this command, is that nearly all pages allocated as shared memory are resident. By having statistics for shared memory pages allocated and resident we can conclude that shared memory too could be allocated in a lazy, alias on demand. Also, there is a difference between resident and allocated, which indicates lazy allocation too.

Inside the database I am aware of two parameters that could influence shared pages usage; pre_page_sga and _touch_sga_pages_during_allocation, see my article on these. However, what is interesting, is that these parameters are different for the instance I am testing with for this blogpost, which is running on a VM:

SYS@testdb AS SYSDBA> @parms
Enter value for parameter: page
old  20: where name like nvl('%&parameter%',name)
new  20: where name like nvl('%page%',name)
Enter value for isset:
old  21: and upper(isset) like upper(nvl('%&isset%',isset))
new  21: and upper(isset) like upper(nvl('%%',isset))
Enter value for show_hidden: Y
old  22: and flag not in (decode('&show_hidden','Y',3,2))
new  22: and flag not in (decode('Y','Y',3,2))

-------------------------------------------------- ---------------------------------------------------------------------- -------- ---------- ----------
olap_page_pool_size    				   0   									  TRUE 	   FALSE      FALSE
pre_page_sga   					   TRUE        								  TRUE 	   FALSE      FALSE
use_large_pages        				   TRUE        								  TRUE 	   FALSE      FALSE
_max_largepage_alloc_time_secs 			   10  									  TRUE 	   FALSE      FALSE
_olap_page_pool_expand_rate    			   20  									  TRUE 	   FALSE      FALSE
_olap_page_pool_hi     				   50  									  TRUE 	   FALSE      FALSE
_olap_page_pool_hit_target     			   100 									  TRUE 	   FALSE      FALSE
_olap_page_pool_low    				   262144      								  TRUE 	   FALSE      FALSE
_olap_page_pool_pressure       			   90  									  TRUE 	   FALSE      FALSE
_olap_page_pool_shrink_rate    			   50  									  TRUE 	   FALSE      FALSE
_realfree_heap_pagesize        			   65536       								  TRUE 	   FALSE      FALSE
_realfree_pq_heap_pagesize     			   65536       								  TRUE 	   FALSE      FALSE
_session_page_extent   				   2048        								  TRUE 	   FALSE      FALSE
_touch_sga_pages_during_allocation     		   FALSE       								  TRUE 	   FALSE      FALSE

14 rows selected.

In the database I created on my VM, pre_page_sga equals to TRUE and _touch_sga_pages_during_allocation to FALSE, which is the exact inverse of the settings of a database (PSU 160419) on a huge machine. Perhaps these parameters are dynamically set based on size of the SGA and logic (if _touch_sga_pages_during_allocation is TRUE, it makes sense to set pre_page_sga to FALSE, as it’s function has been performed by the bequeathing session.

However, having pre_page_sga set to TRUE it makes sense almost all SGA (shared) pages are allocated, because pre_page_sga (at least in Oracle 12, not sure about earlier versions, because the Oracle description of this parameter is different from what happens in Oracle 12) spawns a background process (sa00) that scans SGA pages, which means it pages them, resulting in the actual allocation. Let’s test this by setting pre_page_sga to false, it should lead to way lesser shared memory pages allocated, which will eventually be allocated as database processes are paging them in:

SQL> alter system set pre_page_sga=false scope=spfile;
SQL> startup force;

And then look at ipcs -mu again:

[oracle@oracle-linux [testdb] ~]$ ipcs -mu

------ Shared Memory Status --------
segments allocated 4
pages allocated 256005
pages resident  92696
pages swapped   0
Swap performance: 0 attempts   	 0 successes

As expected, only the bare necessary pages are resident after startup force, all the other shared pages will be slowly paged in as foreground and background processes touching SGA pages during execution.

How would that work when we set sga_max_size to a different value than sga_target? If the pages beyond the sga_target are never allocated, you could control the amount of SGA pages used by setting sga_target, but ‘reserve’ extra memory to use by setting sga_max_size higher, which is never allocated, so it is not wasted. Let’s setup the instance:

SQL> alter system set pre_page_sga=true scope=spfile;
SQL> show spparameter sga_target

SID    	 NAME  			       TYPE    	   VALUE
-------- ----------------------------- ----------- ----------------------------
*      	 sga_target    		       big integer 1000M
SQL> ! ipcs -mu

------ Shared Memory Status --------
segments allocated 4
pages allocated 256005
pages resident  102512
pages swapped   0
Swap performance: 0 attempts   	 0 successes

This sets the pre_page_sga parameter from the spfile, which means the instance will spawn a process to touch SGA pages on next startup.
Currently, the sga_target for sizing the SGA is set to 1000M in the spfile.
ipcs tells us we got 256005 pages are allocated, which makes sense: 256005*4=1024020k, which is slightly more than the set 1000M, which means essentially sga_target equals pages allocated.

SQL> alter system set sga_max_size=2g scope=spfile;
SQL> startup force;
ORACLE instance started.

Total System Global Area 2147483648 bytes
Fixed Size     		    2926472 bytes
Variable Size  		 1358956664 bytes
Database Buffers       	  637534208 bytes
Redo Buffers   		  148066304 bytes
Database mounted.
Database opened.

This sets sga_max_size to double the amount of sga_target, and ‘startup force’ bounces the instance.

SQL> show parameter sga_target

NAME   				     TYPE      	 VALUE
------------------------------------ ----------- ------------------------------
sga_target     			     big integer 1008M

Here we see the actual parameter in the database is set to 1008M. Now let’s look at the ipcs -mu values again:

> !ipcs -mu

------ Shared Memory Status --------
segments allocated 4
pages allocated 524291
pages resident  521923
pages swapped   0
Swap performance: 0 attempts   	 0 successes

521923*4=2087692. So (almost) all the memory set for sga_max_size is allocated. In fact, if you look at the values at instance startup values reported above, you see ‘Total System Global Area’ showing the 2G, it’s all SGA, so it’s all touched because of pre_page_sga being set to TRUE. So the next test would be to have pre_page_sga being set to FALSE:

SQL> alter system set pre_page_sga=false scope=spfile;
SQL> startup force
ORACLE instance started.

Total System Global Area 2147483648 bytes
Fixed Size     		    2926472 bytes
Variable Size  		 1358956664 bytes
Database Buffers       	  637534208 bytes
Redo Buffers   		  148066304 bytes
Database mounted.
Database opened.

All memory is still declared SGA, as we can see. However, by having _touch_sga_pages_during_allocation set to FALSE and pre_page_sga set to FALSE, we should see only the actual used SGA pages being allocated:

SQL> !ipcs -mu

------ Shared Memory Status --------
segments allocated 4
pages allocated 524291
pages resident  91692
pages swapped   0
Swap performance: 0 attempts   	 0 successes

The above output shows the shared memory status directly after I restart my instance, so this is not only less than sga_max_size, it is even less than sga_target (91692*4=336768, ~ 336M). This will grow up to sga_target, because these pages will get paged in by the database processes.

How does this look like when we add in huge pages? In Oracle in my instance the parameter to tell oracle to allocate huge pages if there are any (‘use_large_pages’) is set to TRUE. This will make Oracle use large pages if any are available. This is true, even if there are not enough huge pages to satisfy the entire SGA; Oracle will just allocate all that can be allocated, and create a new shared memory segment using small pages for the remainder of the needed shared memory.

Sadly, it seems per memory segment statistics like rss, pss, shared and private clean and dirty, etc. are not implemented for huge pages:

[oracle@oracle-linux [testdb] ~]$ cat /proc/$(pgrep pmon)/smaps
61000000-d8000000 rw-s 00000000 00:0e 688129                             /SYSV00000000 (deleted)
Size:            1949696 kB
Rss:                   0 kB
Pss:                   0 kB
Shared_Clean:          0 kB
Shared_Dirty:          0 kB
Private_Clean:         0 kB
Private_Dirty:         0 kB
Referenced:            0 kB
Anonymous:             0 kB
AnonHugePages:         0 kB
Swap:                  0 kB
KernelPageSize:     2048 kB
MMUPageSize:        2048 kB
Locked:                0 kB
VmFlags: rd wr sh mr mw me ms de ht sd

This is the main shared memory segment, allocated from huge pages (as can be seen with KernelPageSize and MMUPageSize), which means it’s the segment holding the shared pool and buffercache. This can also be seen by the size: 1949696 kB, which is nearly the 2G of sga_max_size.

However, we can just use the global information on system V shared memory (ipcs -mu) and we can use the huge page information in /proc/meminfo:

[oracle@oracle-linux [testdb] ~]$ grep -i huge /proc/meminfo
AnonHugePages:         0 kB
HugePages_Total:    1100
HugePages_Free:      880
HugePages_Rsvd:      805
HugePages_Surp:        0
Hugepagesize:       2048 kB

The statistics of interest are:
hugepages_total: the total number of huge pages allocated. warning: huge pages memory allocated by the kernel is NOT available for allocation of regular sized pages (which means you can starve your processes and the kernel for normal pages by setting the number of huge pages too high).
hugepages_free: the number of huge pages which are not used currently. warning: this includes allocated but not yet initialised pages, which hugepages_rsvd shows.
hugepages_rsvd: the number of huge pages allocated but not yet initialised.
hugepages_surp: the number of huge pages allocated (truly allocated and not yet initialised) greater than the total number of huge pages set. this value can be greater than zero if the kernel setting vm.nr_overcommit_hugepages is greater than zero. The value of this setting is zero by default, and at least for usage with the Oracle database, this value should remain zero.

The same information can be obtained using ipcs -mu, but with a twist:

[oracle@oracle-linux [testdb] ~]$ ipcs -mu

------ Shared Memory Status --------
segments allocated 4
pages allocated 524803
pages resident  122881
pages swapped   0
Swap performance: 0 attempts   	 0 successes

Some of you might get the twist on this by looking at the number.
It turns out ipcs has no facility for huge pages, it just reports the number of pages as if these were 4 kB.
524803*4 (kB) / 1024 (to make it MB) = 2050.

Now going back to the goal of looking into this: I told shared memory is allocated and paged at startup time when _touch_sga_pages_during_allocation is set to TRUE (set to false as default value in my current database), and it could be explicitly paged by the background process sa00 after startup of the instance when pre_page_sga is set to TRUE. When both are set to false, shared memory allocated from default sized 4kB pages is allocated only when it’s used. In the above examples with huge pages, the tests were done with pre_page_sga set to false. This shows exactly the same ‘lazy allocation’ behaviour as 4kB pages.

When ‘extra’ memory is reserved from the operating system by setting sga_max_size to a higher value than sga_target, this will all be allocated and paged if either _touch_sga_pages_during_allocation or pre_page_sga is set to TRUE, which doesn’t make sense; if the memory is taken, you might as well use it. However, this is different if both _touch_sga_pages_during_allocation and pre_page_sga are set to false. All memory beyond sga_target up to sga_max_size is allocated, but never touched, and thus never paged in, so never truly allocated. Please mind linux itself understands this perfectly (aiming at huge pages and ‘reserved’ pages), however the system V ipc kernel settings do not; you need to set the shared memory values high enough to facilitate the total sum of sga_max_size values, not the truly used sizes as indicated by the sum of sga_target values.

The inspiration for this investigation came from a question on my blog. However, the question was about memory_target and memory_max_target and AIX. I do not have an AIX system at hand. I did not investigate the implementation of memory_target and memory_max_target on AIX. So I can’t comment on that. What I can say, is that on Linux, you really, really should use automatic shared memory management (ASMM) alias setting sga_target or setting it manually (and set huge pages!). If you are used to these memory management settings on databases not on AIX, it probably makes sense to use that on AIX too, even if the automatic memory management (AMM) alias setting memory_target is implemented brilliantly on AIX, for the sake of predictability and standardisation.

In a previous article called ‘memory allocation on startup’ I touched on the subject of NUMA; Non Uniform Memory Access. This article is about how to configure NUMA, how to look into NUMA usage and a real life case of NUMA optimisation using in-memory parallel execution.

At this point in time (start of the summer of 2016) we see that the CPU speed competition has stagnated and settled at somewhere below maximally 4 gigahertz, and instead the number of core’s and the size of memory is growing. The common used server type in the market I am in is a two socket server. It is not unthinkable that in the near future servers with more than two sockets will increase in popularity, which facilitates the increase in (parallel) computing capacity and maximal amount of memory.

On the other hand of course there are already 2+ socket count servers, of which some have been multi socket for quite a while: Oracle/Sun M series, IBM P series, and a large diversity of intel based 2+ socket servers, like the Oracle/Sun “-4” and “-8” series. The investigations in this blogpost are done done on a SGI UV 300RL, which is a system with (up to) 32 sockets, 576 cores, and 1152 CPU threads.

There already have been a lot of blogposts about Linux and NUMA:
Kevin Closson
Martin Bach: linux-large-pages-and-non-uniform-memory-distribution, _enable_numa_support and numactl
Yves Colin
Bertrand Drouvot
I am sure there will be more, these are the ones that I read in the past.

This article will take you through the configuration of NUMA for the Oracle database and how to measure specifics from the Linux side. It will touch some of the things mentioned in the other blogposts, in order to provide a consistent story.
The database version is:
The operating system is Oracle Linux 7.1
Kernel: 3.8.13-98.4.1.el7uek.x86_64

1. NUMA or not.
In order to sensibly use NUMA optimisation with the Oracle database, you first need to have hardware that is actually NUMA capable. This means multiple sockets in use that have local memory (please mind this includes common two socket server hardware). The next thing needed to use NUMA is NUMA support enabled by the Linux kernel. With the current version of the oracle-preinstall rpm, NUMA is disabled on the kernel load line in grub (numa=off). When NUMA is disabled, it will show one NUMA node (=socket) with numactl.

First make sure you are on NUMA enabled hardware by issuing ‘numactl –hardware’ and see multiple nodes specified.

2. Default Oracle.
When you don’t modify (undocumented) NUMA parameters, Oracle will NOT run in NUMA aware mode. In other words, you explicitly need to enable NUMA in the database. This is also visible if you look in the numa_maps, which shows process memory and NUMA information:

$ grep huge /proc/$(pgrep pmon_slob)/numa_maps
60000000 interleave:0-31 file=/SYSV00000000\040(deleted) huge dirty=2 mapmax=904 N0=1 N1=1
80000000 interleave:0-31 file=/SYSV00000000\040(deleted) huge dirty=3339 mapmax=904 N0=112 N1=98 N2=98 N3=98 N4=98 N5=98 N6=98 N7=97 N8=96 N9=96 N10=96 N11=101 N12=108 N13=106 N14=105 N15=105 N16=107 N17=107 N18=106 N19=109 N20=106 N21=106 N22=107 N23=107 N24=106 N25=107 N26=108 N27=108 N28=109 N29=112 N30=111 N31=113
9c460000000 interleave:0-31 file=/SYSV00000000\040(deleted) huge
9c480000000 prefer:1 file=/SYSV85a066e4\040(deleted) huge dirty=1 mapmax=904 N0=1

Few things to note: I took the pmon process, and my instance name is slob (pmon_slob). The SGA only is allocated from huge pages, so I grepped for ‘huge’. numa_maps only shows the process’ paged memory, my SGA is way bigger. The important thing to see here is the main SGA memory allocation at 80000000 reads ‘interleave:0-31’, which means the memory is spread over all the NUMA nodes. This can also be seen on the same line, because the memory allocation per node is shown with Nnn=nr.pages.

3. Set parameters
In Oracle you need to set two parameters to fully enable NUMA support in the database:

alter system set "_enable_NUMA_support"=true scope=spfile;
alter system set "_px_numa_support_enabled"=true scope=spfile;

These are undocumented parameters. However, these parameters are explicitly mentioned in MOS article ‘What Is Correct Setting for _px_numa_support_enabled for 12c Databases That Wish to Use NUMA?’ (Doc ID 1956463.1).

Once you have set these parameters, Oracle will start up in NUMA aware mode. One of the things that is different in NUMA aware mode, is how the buffercache is created. This can be seen by looking at the numa_maps file of pmon again:

$ grep huge /proc/$(pgrep pmon_slob)/numa_maps
60000000 interleave:0-31 file=/SYSV00000000\040(deleted) huge dirty=2 mapmax=904 N0=1 N1=1
80000000 interleave:0-31 file=/SYSV00000000\040(deleted) huge dirty=3587 mapmax=904 N0=112 N1=112 N2=112 N3=112 N4=112 N5=112 N6=112 N7=112 N8=112 N9=112 N10=112 N11=112 N12=113 N13=112 N14=113 N15=112 N16=112 N17=112 N18=113 N19=112 N20=112 N21=112 N22=112 N23=112 N24=112 N25=112 N26=112 N27=112 N28=112 N29=112 N30=112 N31=112
8860000000 prefer:25 file=/SYSV00000000\040(deleted) huge dirty=1536 N25=1536
d260000000 prefer:26 file=/SYSV00000000\040(deleted) huge dirty=512 N26=512
11c40000000 prefer:27 file=/SYSV00000000\040(deleted) huge dirty=512 N27=512
16620000000 prefer:28 file=/SYSV00000000\040(deleted) huge dirty=512 N28=512
1b000000000 prefer:29 file=/SYSV00000000\040(deleted) huge dirty=512 N29=512
1f9e0000000 prefer:30 file=/SYSV00000000\040(deleted) huge dirty=256 N30=256
243c0000000 prefer:31 file=/SYSV00000000\040(deleted) huge dirty=512 N31=512
28da0000000 prefer:0 file=/SYSV00000000\040(deleted) huge dirty=256 N0=256
2d780000000 prefer:1 file=/SYSV00000000\040(deleted) huge dirty=512 N1=512
32160000000 prefer:2 file=/SYSV00000000\040(deleted) huge dirty=256 N2=256
36b40000000 prefer:3 file=/SYSV00000000\040(deleted) huge dirty=512 N3=512
3b520000000 prefer:4 file=/SYSV00000000\040(deleted) huge dirty=256 N4=256
3ff00000000 prefer:5 file=/SYSV00000000\040(deleted) huge dirty=512 N5=512
448e0000000 prefer:6 file=/SYSV00000000\040(deleted) huge dirty=256 N6=256
492c0000000 prefer:7 file=/SYSV00000000\040(deleted) huge dirty=512 N7=512
4dca0000000 prefer:8 file=/SYSV00000000\040(deleted) huge dirty=256 N8=256
52680000000 prefer:9 file=/SYSV00000000\040(deleted) huge dirty=512 N9=512
57060000000 prefer:10 file=/SYSV00000000\040(deleted) huge dirty=256 N10=256
5ba40000000 prefer:11 file=/SYSV00000000\040(deleted) huge dirty=512 N11=512
60420000000 prefer:12 file=/SYSV00000000\040(deleted) huge dirty=256 N12=256
64e00000000 prefer:13 file=/SYSV00000000\040(deleted) huge dirty=512 N13=512
697e0000000 prefer:14 file=/SYSV00000000\040(deleted) huge dirty=256 N14=256
6e1c0000000 prefer:15 file=/SYSV00000000\040(deleted) huge dirty=512 N15=512
72ba0000000 prefer:16 file=/SYSV00000000\040(deleted) huge dirty=256 N16=256
77580000000 prefer:17 file=/SYSV00000000\040(deleted) huge dirty=512 N17=512
7bf60000000 prefer:18 file=/SYSV00000000\040(deleted) huge dirty=256 N18=256
80940000000 prefer:19 file=/SYSV00000000\040(deleted) huge dirty=512 N19=512
85320000000 prefer:20 file=/SYSV00000000\040(deleted) huge dirty=256 N20=256
89d00000000 prefer:21 file=/SYSV00000000\040(deleted) huge dirty=512 N21=512
8e6e0000000 prefer:22 file=/SYSV00000000\040(deleted) huge dirty=256 N22=256
930c0000000 prefer:23 file=/SYSV00000000\040(deleted) huge dirty=512 N23=512
97aa0000000 prefer:24 file=/SYSV00000000\040(deleted) huge dirty=512 N24=512
9c460000000 prefer:26 file=/SYSV00000000\040(deleted) huge
9c480000000 prefer:2 file=/SYSV85a066e4\040(deleted) huge dirty=1 mapmax=904 N5=1

If we look at the output, a few segments are still allocated interleaved (these are the fixed SGA and the shared pool), and we see allocations for every NUMA node. These per NUMA node allocations are the buffer cache.

Okay. Now the database is setup for NUMA, and we checked how this looks like on the operating system level using Linux tools. However, we need to realise that NUMA is no magic that sets your database into the next gear. Most things you do are not NUMA optimised by the database, simply because it wouldn’t make sense.

For example, when I perform a buffered full table scan without parallel query, it will spread out the blocks over the allocated segments. Why? Because my session is running on a random CPU thread which could run on another CPU thread the next time it is scheduled. What I am saying here is: only specific things are NUMA optimised.

A next logical question would be: but then what is actually NUMA optimised? One obvious thing that is NUMA optimised is the database writer (this is mentioned in the alert.log at startup). Another NUMA optimised thing is parallel query. If you look at the parallel query worker processes, the status file gives a hint:

# grep allowed_list /proc/$(pgrep p000_slob)/status
Cpus_allowed_list:	450-467,1026-1043
Mems_allowed_list:	0-31

This shows that the first parallel query worker process is limited in what CPU (threads) it is allowed to use. All parallel query worker processes are limited to ranges of CPU threads. Please mind it is NOT limited in what memory it can use, so it can use both local and remote memory. It may look weird that it still can access remote memory, totally ignoring NUMA optimised (=local) memory access.

First the CPU thread ranges. If you look what the range mean (it’s rather obvious), you see this:

# lscpu | grep 450
NUMA node25 CPU(s):    450-467,1026-1043

The CPU thread range is exactly a NUMA node. So whatever happens, parallel query worker process p000 can only run on the CPU threads of NUMA node 25!

Then absence of limits in memory: this means the parallel query process can still access anything it needs. This means that anything that used to work with parallel query can still run as it did without or before the NUMA optimisations. In specific cases, parallel query can take advantage of the NUMA setting.

One of such cases is in-memory parallel execution. With in-memory parallel execution, the parallel query slaves read blocks in the buffer cache instead of performing a direct path read. In-memory parallel execution is enabled when the parameter parallel_degree_policy is set to auto, and heuristically determines feasibility.

I setup a table of approximately 700GB, and performed an in-memory parallel query scan on a freshly started instance, in order to have the parallel query worker processes have as little memory touched as possible. At the end of the scan, this is how numa_maps of parallel query worker process p000 looks like:

# grep huge /proc/$(pgrep p000_slob)/numa_maps
60000000 interleave:0-31 file=/SYSV00000000\040(deleted) huge dirty=2 mapmax=906 N0=1 N1=1
80000000 interleave:0-31 file=/SYSV00000000\040(deleted) huge dirty=4608 N0=144 N1=144 N2=144 N3=144 N4=144 N5=144 N6=144 N7=144 N8=144 N9=144 N10=144 N11=144 N12=144 N13=144 N14=144 N15=144 N16=144 N17=144 N18=144 N19=144 N20=144 N21=144 N22=144 N23=144 N24=144 N25=144 N26=144 N27=144 N28=144 N29=144 N30=144 N31=144
8880000000 prefer:25 file=/SYSV00000000\040(deleted) huge dirty=50176 N25=50176
d2a0000000 prefer:26 file=/SYSV00000000\040(deleted) huge dirty=6145 mapmax=70 N26=6145
11c80000000 prefer:27 file=/SYSV00000000\040(deleted) huge dirty=4608 N27=4608
16660000000 prefer:28 file=/SYSV00000000\040(deleted) huge dirty=5121 mapmax=70 N28=5121
1b040000000 prefer:29 file=/SYSV00000000\040(deleted) huge dirty=14336 N29=14336
1fa20000000 prefer:30 file=/SYSV00000000\040(deleted) huge dirty=5121 mapmax=72 N30=5121
24400000000 prefer:31 file=/SYSV00000000\040(deleted) huge dirty=4608 N31=4608
28de0000000 prefer:0 file=/SYSV00000000\040(deleted) huge dirty=6145 mapmax=70 N0=6145
2d7c0000000 prefer:1 file=/SYSV00000000\040(deleted) huge dirty=4099 N1=4099
321a0000000 prefer:2 file=/SYSV00000000\040(deleted) huge dirty=2528 mapmax=73 N2=2528
36b80000000 prefer:3 file=/SYSV00000000\040(deleted) huge dirty=2222 N3=2222
3b560000000 prefer:4 file=/SYSV00000000\040(deleted) huge dirty=2504 mapmax=69 N4=2504
3ff40000000 prefer:5 file=/SYSV00000000\040(deleted) huge dirty=2249 N5=2249
44920000000 prefer:6 file=/SYSV00000000\040(deleted) huge dirty=4754 mapmax=69 N6=4754
49300000000 prefer:7 file=/SYSV00000000\040(deleted) huge dirty=2296 N7=2296
4dce0000000 prefer:8 file=/SYSV00000000\040(deleted) huge dirty=2463 mapmax=70 N8=2463
526c0000000 prefer:9 file=/SYSV00000000\040(deleted) huge dirty=2246 N9=2246
570a0000000 prefer:10 file=/SYSV00000000\040(deleted) huge dirty=2440 mapmax=72 N10=2440
5ba80000000 prefer:11 file=/SYSV00000000\040(deleted) huge dirty=2233 N11=2233
60460000000 prefer:12 file=/SYSV00000000\040(deleted) huge dirty=2469 mapmax=69 N12=2469
64e40000000 prefer:13 file=/SYSV00000000\040(deleted) huge dirty=2241 N13=2241
69820000000 prefer:14 file=/SYSV00000000\040(deleted) huge dirty=2434 mapmax=69 N14=2434
6e200000000 prefer:15 file=/SYSV00000000\040(deleted) huge dirty=2342 N15=2342
72be0000000 prefer:16 file=/SYSV00000000\040(deleted) huge dirty=2461 mapmax=70 N16=2461
775c0000000 prefer:17 file=/SYSV00000000\040(deleted) huge dirty=2329 N17=2329
7bfa0000000 prefer:18 file=/SYSV00000000\040(deleted) huge dirty=2499 mapmax=71 N18=2499
80980000000 prefer:19 file=/SYSV00000000\040(deleted) huge dirty=2334 N19=2334
85360000000 prefer:20 file=/SYSV00000000\040(deleted) huge dirty=2466 mapmax=73 N20=2466
89d40000000 prefer:21 file=/SYSV00000000\040(deleted) huge dirty=2328 N21=2328
8e720000000 prefer:22 file=/SYSV00000000\040(deleted) huge dirty=2180 mapmax=69 N22=2180
930e0000000 prefer:23 file=/SYSV00000000\040(deleted) huge dirty=2208 mapmax=69 N23=2208
97aa0000000 prefer:24 file=/SYSV00000000\040(deleted) huge dirty=2167 mapmax=68 N24=2167
9c460000000 prefer:26 file=/SYSV00000000\040(deleted) huge dirty=2 mapmax=147 N26=2
9c480000000 prefer:0 file=/SYSV85a066e4\040(deleted) huge dirty=1 mapmax=906 N2=1

If you look at memory address 8880000000, you see parallel query worker process p000 used 50176 pages, while on the other (buffer cache/’prefer’) segments it touched somewhere between 2200-6150 pages. So it clearly used it’s local memory more than the remote memories. This is proof in-memory parallel query is NUMA optimised.

In fact, when I time the query using in-memory parallel query with NUMA enabled, it takes a little less than 8 seconds consistently once the data is loaded in the buffercache. When I disable NUMA optimisations (_enable_NUMA_support=false), and execute the same query, also doing in-memory parallel query, the time goes up to a little less than 24 seconds, only scanning the buffercache.

%d bloggers like this: