Applies to:
Oracle Server - Enterprise Edition - Version 10.2.0.3 and laterInformation in this document applies to any platform.
***Checked for relevance on 27-Apr-2012***
Goal
Migrated a database server from a server with 8 CPU's to a server with 128 (virtual) CPU's - that is, 2 CPU's, each with 8 cores which each handled 8 threads = 2x8x8 = 128.So Solaris (on SPARC T series) is presenting Oracle with 128 CPU's. This makes Oracle adjust it's memory demands upwards, meaning that before the databases can ran fine with 320 MB of ram (sga_target) now needs to have up around 800-900 MB of RAM?
Is there anyway to minimize the memory usage?
Fix
Calculating the sga_target should consider the following:The calculation for min cache size = max (number of CPUs, number of processor groups) * max (granule size, 4MB)
So the SGA size will mainly depend on CPU_COUNT.
Bug 8316255 was filed for this issue and initially closed as 'Not a Bug' with the comments:
"CPU_COUNT should not be changed . This is not a bug.
Reducing CPU_COUNT has side effects due to the fact that the per cpu power of CMT may not be sufficient to sustain high load ( e.g. cache fusion messages ).
Also, some of the dependent parameters such as gc latches would be underconfigured ".
Currently, a fix has been produced which added the notion of the raw cpu threads and effective cpu threads.
This fix is only available in the 11.2.0.3 Bundle Patch 1 for Exadata Database and will be included in 12.1.
Please refer to:
Note 8316255.8 - Bug 8316255 - Wrong discovery of number of CPUs in SPARC T series.
Note 13343057.8 - Bug 13343057 - 11.2.0.3 Bundle Patch 1 for Exadata Database
For other systems, a way of reducing memory needs would be to set a manual value for DB_CACHE_SIZE (which depends on CPU_COUNT) and keep the old value of SGA_TARGET or not using ASMM (not using SGA_TARGET).
No comments:
Post a Comment