Thursday 8 September 2016

memlock configuration on linux for 12c part I

Going through the Oracle 12c Installation docs for linux, I came across the following memlock requirement:


So what is 'memlock' and where this requirements are coming from?

Lets see how it works, I have 1GB (for a reason) physical memory configured and SGA_TARGET=900M on 12.1.0.2

1. Lets configure Hugepages and set memlock much lower then Hugepages configured and see whats going to happen.

[oracle@single1 ~]$ cat /etc/sysctl.conf |  grep vm
vm.nr_hugepages = 454

[oracle@single1 ~]$ ulimit -l
217728

[root@single1 ~]# grep Huge /proc/meminfo
HugePages_Total:     447
HugePages_Free:      447
HugePages_Rsvd:        0
HugePages_Surp:        0
Hugepagesize:       2048 kB

Upon the instance startup, oracle provided the following in the alertlog

Starting ORACLE instance (normal) (OS id: 2308)
Thu Sep 08 00:52:56 2016
CLI notifier numLatches:3 maxDescs:519
Thu Sep 08 00:52:56 2016
**********************************************************************
Thu Sep 08 00:52:56 2016
Dump of system resources acquired for SHARED GLOBAL AREA (SGA)

Thu Sep 08 00:52:56 2016
 Per process system memlock (soft) limit = 213M
Thu Sep 08 00:52:56 2016
 Expected per process system memlock (soft) limit to lock
 SHARED GLOBAL AREA (SGA) into memory: 902M
Thu Sep 08 00:52:56 2016
 Available system pagesizes:
  4K, 2048K
Thu Sep 08 00:52:56 2016
 Supported system pagesize(s):
Thu Sep 08 00:52:56 2016
  PAGESIZE  AVAILABLE_PAGES  EXPECTED_PAGES  ALLOCATED_PAGES  ERROR(s)
Thu Sep 08 00:52:56 2016
        4K       Configured               5          228357        NONE
Thu Sep 08 00:52:56 2016
     2048K              447             451               5        NONE
Thu Sep 08 00:52:56 2016
RECOMMENDATION:
Thu Sep 08 00:52:56 2016
 1. For optimal performance, configure system with expected number
 of pages for every supported system pagesize prior to the next
 instance restart operation.
Thu Sep 08 00:52:56 2016
 2. Increase per process memlock (soft) limit to at least 902MB
 to lock 100% of SHARED GLOBAL AREA (SGA) pages into physical memory
Thu Sep 08 00:52:56 2016
**********************************************************************


[root@single1 ~]# grep Huge /proc/meminfo
HugePages_Total:     447
HugePages_Free:      442
HugePages_Rsvd:        0
HugePages_Surp:        0
Hugepagesize:       2048 kB

What we see now that only 5 of Hugepages are used and system is experiencing heavy trashing. The instance has never opened and I aborted it after 1 h of wait.

[oracle@single1 ED12]$ vmstat 1
procs -----------memory---------- ---swap-- -----io---- --system-- -----cpu-----
 r  b   swpd   free   buff  cache   si   so    bi    bo   in   cs us sy id wa st
 9  5 327268   8276     76   2548 25329 5873 111710  5943 7419 3964  1 81  9 10  0
 8  9 327192   8304     76   2780 31724 4704 147192  4712 7993 4184  1 86  0 13  0
12  6 327416   8208     76   2604 35900 5240 150804  5240 9309 4434  0 98  0  2  0
13  6 327312   8092     76   2876 43788 5624 120172  5640 9220 3972  1 94  0  5  0
 9  5 327156   8268     76   3188 41672 5416 144988  5416 9602 4268  0 96  0  4  0
11  7 326944   8332     76   2856 42608 6332 132128  6356 9086 4061  1 94  0  5  0
14  8 327288   8072     76   2880 40352 5416 114864  5436 9221 4477  0 75  0 25  0
10  4 327464   8128     76   2608 40860 5144 125752  5240 8501 4199  1 78  0 21  0
10  7 327468   8092     80   2652 38696 3784 137976  3828 8758 3978  0 98  0  2  0
11  9 327364   8224     80   2816 40768 3692 138048  3704 8603 3874  0 91  0  9  0
12  5 327492   8472     88   2620 45792 4200 146660  4244 9484 4267  1 94  0  5  0
 8  4 327396   8264     88   2896 42648 5616 124336  5616 9898 4251  1 96  0  3  0
11  5 327352   8116     84   2752 35832 5616 124496  5640 8382 4228  0 92  0  8  0
10  7 327272   8148     88   2700 35352 6628 117264  6628 8418 4152  1 96  0  3  0


This simple experiment shows that:

1. Setting the memlock incorrectly makes all Hugepages not to be utilised by the database. This is just a waist of memory, no other process are able to use them. In case of memory shortage this could lead to the trashing of the system (as in our case, where its only 1GB of memory and 900MB of are Hugepages)

2. Also starting from 11.2.0.3 oracle should use all available Hugepages even if number of hugepaes configured are less then shole SGA. It looks like it is required that all configured hugepages should be able to be locked into memory.

3. Therefore, the proper memlock configuration is very important  and should be correlated to the number of hugepages and not to the total amount of physical memory.

No comments:

Post a Comment