ipcs 之后查看共享内存段，发现有oracle三条记录

sundog315

重启后ipcs还是三块？

tolilong

sundog315 发表于 2012-11-12 15:18
重启后ipcs还是三块？

还是三块.

zhangdberic

我也是三快共享段，有人解决这个问题了吗

Yong Huang

I checked a few servers I have. It looks like something new in 11g (or 12c). My 10g server has only 1 segment. 11g has 3. One 2 MB, one large segment for the most part of my HugePages (from which almost 99% is allocated to SGA), and one small one of varying size (16M to 128M). According to /proc/<pid>/smaps (available in newer version of Linux), the first piece is the small but varying size segment, then the big one, then the fixed size 2M segment. Not sure why Oracle does that.

bluemoon0083

Yong Huang 发表于 2013-8-20 06:37
I checked a few servers I have. It looks like something new in 11g (or 12c). My 10g server has only  ...

http://www.itpub.net/thread-1810326-1-4.html

Why Multiple Shared Memory Segments are Created From 11.2.0.3 (Doc ID 1399908.1)

Goal
This document explains why databases have multiple shared segments after upgrading to 11.2.0.3, without making any change in the settings and parameters.


Solution
From 11.2.0.3, all databases are expected to have multiple shared memory segments.

The additional shared memory segments are an expected side effect of the fix for unpublished Bug 12654172 included 11.2.0.3. Part of that fix is to ensure that each shared memory segment making up the SGA only contains sub-areas with an identical alignment requirement - hence we end up spreading the SGA over more separate SHM segments..

In cases of very large numbers of databases this change may require the customer to review of the OS parameter SHMMNI (or equivalent), but otherwise this change should not be a problem - we have always allowed the SGA to split over multiple SHM segments - this change just adds an additional reason that can cause the split.

Yong Huang

Thanks for the MOS note. The unpublished bug 12654172 is the base bug for these two bugs:

Bug 10155276 : WHEN FORCING HUGEPAGES STARTUP SHOULD ABORT ON FIRST SHMGET FAILURE
Bug 12725442 : INCORRECT HUGEPAGES AND INSTANCE MEMORY PARAMETERS CAUSE DBNODE PANICS

Both are specific to the Linux port of Oracle. The second one is quite useless. The first one basically says before 11.2.0.3, if you have enabled NUMA support (which is enabled by default before 11.2.0.1 but disabled beginning with 11.2.0.1), and you have over-allocated total SGA that can use HugePages and use_large_pages=ONLY, then HugePages are still allocated (and possibly instances are started), while the correct behavior is that the instances should fail to start. Splitting the one monolithic shared memory segment into 3 should solve the problem, but I don't know how it does. To "ensure that each shared memory segment making up the SGA only contains sub-areas with an identical alignment requirement" is the key. I wish there was more info about it.

I tested on an 8-CPU Linux box running Oracle 11.2.0.3 by changing _enable_NUMA_support to true. After startup, I see 6 shared memory segments instead of 3. (numactl tells me I have 2 NUMA nodes.)

xieyunzheng

我也有同样的问题，能解释一下吗？

sqysl

楼上不是给出解释了吗？