inmemory: sa00 process

After the restart of a 12c inmemory database with 300GB+ SGA, I noticed that an Oracle background process sa00 was consuming a bit of CPU. Documentation suggests that it is SGA Allocator process, however, ipcs -ma command shows that the shared memory segment is already allocated. I was curious, of course, what would that background process will be allocating?.

pstack

Process stack of the process shows that it is touching SGA pages to pre-page SGA memory pages.

$ pstack 21131
#0  0x0000000000d9996e in ksmprepage_memory ()
#1  0x0000000000d99369 in ksm_prepage_sga_seg ()
#2  0x0000000003a5c78b in skgmapply ()
#3  0x0000000000da686a in ksmapply_v2 ()
#4  0x0000000000d9a82c in ksmprepage ()
#5  0x0000000000d99f89 in ksm_sslv_exec_cbk ()
#6  0x0000000000f79810 in ksvrdp ()
#7  0x00000000031013b7 in opirip ()
#8  0x0000000001bb0a08 in opidrv ()
#9  0x00000000026c0f71 in sou2o ()
#10 0x0000000000bbd85e in opimai_real ()
#11 0x00000000026cb6bc in ssthrdmain ()
#12 0x0000000000bbd72c in main ()

$ ps -ef|grep  21131
oracle   21131     1 96 15:00 ?        00:01:50 ora_sa00_XXXXXX

Two notable changes in this area:
1. Incidentally, pre_page_sga initialization parameter was defaulted to a value of FALSE until 11.2. In version 12.1, the parameter value defaults to TRUE.
2. As huge SGA is expected for inmemory databases, a new background process SA00 is also created to touch all SGA memory pages at startup.

As inmemory worker processes will be populating the inmemory column store soon after the startup, touching memory pages at instance startup makes sense, and the feature should improve the performance of inmemory population. At least, Worker processes doesn’t need to suffer from huge amount page faults. (note that this SGA is not using hugepages).

Change to the parameter pre_page_sga also should improve the performance of inmemory scan, as the memory map entries will be setup at the process startup. However, I am not quite clear, how this change will affect the performance of a connection storm, i.e. if there are numerous database connections in a short period of time and disconnects. Isn’t that the reason why the pre_page_sga was defaulted to FALSE? But, I need to test this thoroughly to understand the implications further.

inmemory area is another sub-heap of the top-level SGA heap

I blogged earlier about heap dump shared pool heap duration and was curious to see how the inmemory – 12.1.0.2 new feature – is implemented. This is a short blog entry to discuss the inmemory area heap.

Parameters

I have set the initialization parameters sga_target=32G and inmemory_size=16G, meaning, out of 32GB SGA, 16GB will be allocated to inmemory area and the remaining 16GB will be allocated to the traditional areas such as buffer_cache, shared_pool etc. I was expecting v$sgastat view to show the memory allocated for inmemory area, unfortunately, there are no rows marked for inmemory area (Command “show sga” shows the inmemory area though). However, dumping heapdump at level 2 shows that the inmemory area is defined as a sub-heap of the top level SGA heap. Following are the commands to take an heap dump.

oradebug setmypid
oradebug heapdump 2 -- this command creates an heap dump trace file.
oradebug tracefile_name

Reviewing trace file

Trace file shows that the inmemory area is implemented as two sub-heaps namely IMCA_RO and IMCA_RW. Split is not equal between these two sub-heaps and I am not exactly sure about the algorithm for this split, about 12.75GB is allocated for IMCA_RO and the remaining 3.25GB is allocated for IMCA_RW area [ That’s about 80-20:) split ].

$ grep "heap name" *ora_56235*.trc
HEAP DUMP heap name="sga heap"  desc=0x600013d0
HEAP DUMP heap name="sga heap(1,0)"  desc=0x60063740
HEAP DUMP heap name="sga heap(1,3)"  desc=0x60068048
HEAP DUMP heap name="sga heap(2,0)"  desc=0x6006d490
HEAP DUMP heap name="sga heap(2,3)"  desc=0x60071d98
HEAP DUMP heap name="sga heap(3,0)"  desc=0x600771e0
...
HEAP DUMP heap name="sga heap(7,0)"  desc=0x6009e720
HEAP DUMP heap name="sga heap(7,3)"  desc=0x600a3028
HEAP DUMP heap name="IMCA_RO"  desc=0x60001130 <--- In memory Read only area?
HEAP DUMP heap name="IMCA_RW"  desc=0x60001278 <--- In memory Read write area?

You can learn all about SGA heap duration here , only last two lines are interesting to this blog entry and shows that two sub-heaps were allocated for Inmemory area.

The inmemory sub-heaps are split in to memory extents, similar to traditional SGA heap allocations. Each extent has numerous 64MB chunks allocated to it. These chunks are tagged as “cimadrv”. Total heap size is about 12.5GB.

HEAP DUMP heap name="IMCA_RO"  desc=0x60001130
 extent sz=0x1040 alt=288 het=32767 rec=0 flg=2 opc=2
 parent=(nil) owner=(nil) nex=(nil) xsz=0x30600000 heap=(nil)
 fl2=0x20, nex=(nil), dsxvers=1, dsxflg=0x0
 dsx first ext=0x64000000
 dsx empty ext bytes=0  subheap rc link=0x64000070,0x64000070
 pdb id=0
EXTENT 0 addr=0x363a00000
  Chunk        363a00010 sz=  8388304    free      "               "
  Chunk        3641ffee0 sz= 65011736    freeable  "cimadrv        "
  Chunk        367fffef8 sz= 67108888    freeable  "cimadrv        " <-- 64MB chunks
  Chunk        36bffff10 sz= 67108888    freeable  "cimadrv        "
  Chunk        36fffff28 sz= 67108888    freeable  "cimadrv        "
  Chunk        373ffff40 sz= 67108888    freeable  "cimadrv        "
...
EXTENT 1 addr=0x2e3b00000
  Chunk        2e3b00010 sz= 66059528    freeable  "cimadrv        "
  Chunk        2e79ffd18 sz= 67108888    freeable  "cimadrv        "
  Chunk        2eb9ffd30 sz= 67108888    freeable  "cimadrv        "
…
Total heap size    =13690208144 <-- Total heap size.

Next heap IMCA_RW is more interesting. This sub-heap also has extents with 64MB of chunks allocated it, however, I see that there are also smaller chunks in the heap. (I am still researching meaning of these chunks and trying to avoid guess work at this time.)

EAP DUMP heap name="IMCA_RW"  desc=0x60001278
 extent sz=0x1040 alt=304 het=32767 rec=0 flg=2 opc=2
 parent=(nil) owner=(nil) nex=(nil) xsz=0x50100000 heap=(nil)
 fl2=0x20, nex=(nil), dsxvers=1, dsxflg=0x0
 dsx first ext=0x790000030
 dsx empty ext bytes=0  subheap rc link=0x7900000a0,0x7900000a0
 pdb id=0
EXTENT 0 addr=0x80ff00000
  Chunk        80ff00010 sz= 17825296    free      "               "
  Chunk        810fffe20 sz= 50331672    freeable  "cimadrv        "
  Chunk        813fffe38 sz= 67108888    freeable  "cimadrv        "
  Chunk        817fffe50 sz= 67108888    freeable  "cimadrv        "
…
  Chunk        80f8d5ef8 sz=     8296    freeable  "cimcadrv-sb    " <-- smaller chunks. Most are about 8k or 16k.
  Chunk        80f8d7f60 sz=       48    freeable  "cimcadrv-sbrcv "
  Chunk        80f8d7f90 sz=      184    freeable  "cimcadrv-sblatc"
  Chunk        80f8d8048 sz=     8296    freeable  "cimcadrv-sb    "
  Chunk        80f8da0b0 sz=       48    freeable  "cimcadrv-sbrcv "
  Chunk        80f8da0e0 sz=      184    freeable  "cimcadrv-sblatc"
…
Total heap size    =3489660848

So, if this is similar to shared pool heap, is it possible to get an out-of-space error such as ORA-4031 for the shared pool heap?. There is such an error associated with inmemory option :).

 oerr ora 64356
64356, 00000, "in-memory area out of space"
// *Document: NO
// *Cause:    The in-memory area had no free space.
// *Action:   Drop the in-memory segments to make space.

In summary, I was expecting inmemory area to be allocated as integral part of buffer_cache buffers, however, that is not the case. Inmemory area size is allocated as sub-heaps very similar to the shared pool sub-heaps (but, NOT part of shared pool heaps though). As the software was released just recently, I need to research further to understand the intricate details.