[
https://issues.apache.org/jira/browse/TRAFODION-2733?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16163925#comment-16163925
]
ASF GitHub Bot commented on TRAFODION-2733:
-------------------------------------------
Github user selvaganesang commented on a diff in the pull request:
https://github.com/apache/incubator-trafodion/pull/1228#discussion_r138498956
--- Diff: core/sql/optimizer/RelExpr.cpp ---
@@ -1988,77 +1988,62 @@ NABoolean RelExpr::containsNode(OperatorTypeEnum
nodeType)
}
double RelExpr::computeMemoryQuota(NABoolean inMaster,
- NABoolean perCPU,
- double BMOsMemoryLimit, // in bytes
- UInt16 totalNumBMOs, // per CPU
- double totalBMOsMemoryUsage, // per
CPU, in bytes
+ NABoolean perNode,
+ double BMOsMemoryLimit, // in MB
+ UInt16 totalNumBMOs, // per query
+ double totalBMOsMemoryUsage, // for all
BMOs per node in bytes
UInt16 numBMOsPerFragment, // per
fragment
- double BMOsMemoryUsagePerFragment //
per fragment, in bytes
+ double bmoMemoryUsage, // for the
current BMO/Operator per node in bytes
+ Lng32 numStreams,
+ double &bmoQuotaRatio
)
{
- if ( perCPU == TRUE ) {
- Lng32 exeMem = Lng32(BMOsMemoryLimit/(1024*1024));
+ if ( perNode == TRUE ) {
+ Lng32 exeMem = Lng32(BMOsMemoryLimit/(1024*1024));
- if ( inMaster && CmpCommon::getDefault(ODBC_PROCESS) == DF_ON ) {
-
- // Limiting the total memory in the master process when in both
- // the per-CPU estimation and the ODBC mode.
-
- NADefaults &defs = ActiveSchemaDB()->getDefaults();
-
- Lng32 inCpuLimitDelta =
- defs.getAsLong(EXE_MEMORY_AVAILABLE_IN_MB)
- -
- defs.getAsLong(EXE_MEMORY_RESERVED_FOR_MXOSRVR_IN_MB);
-
- if ( inCpuLimitDelta < 0 )
- inCpuLimitDelta = 50;
-
- if (exeMem > inCpuLimitDelta)
- exeMem = inCpuLimitDelta;
+ // the quota is allocated in 2 parts
+ // The constant part divided equally across all bmo operators
+ // The variable part allocated in proportion of the given BMO operator
+ // estimated memory usage to the total estimated memory usage of all
BMOs
+
+ // The ratio can be capped by the CQD
+ double equalQuotaShareRatio = 0;
+ equalQuotaShareRatio =
ActiveSchemaDB()->getDefaults().getAsDouble(BMO_MEMORY_EQUAL_QUOTA_SHARE_RATIO);
+/*
+ char *equalQuotaShareRatioStr =
getenv("BMO_MEMORY_EQUAL_QUOTA_SHARE_RATIO");
+ if (equalQuotaShareRatioStr != NULL)
+ equalQuotaShareRatio = atof(equalQuotaShareRatioStr);
+*/
+ double constMemQuota = 0;
+ double variableMemLimit = exeMem;
+ if (equalQuotaShareRatio > 0 && totalNumBMOs > 1) {
+ constMemQuota = (exeMem * equalQuotaShareRatio )/ totalNumBMOs;
+ variableMemLimit = (1-equalQuotaShareRatio) * exeMem;
}
-
- // the quota is propotional to both the # of BMOs and the estimated
memory
- // usage in the fragment, and evenly distrbuted among BMOs in the
fragment.
- return ((exeMem/2) * (BMOsMemoryUsagePerFragment/totalBMOsMemoryUsage
+
- double(numBMOsPerFragment)/totalNumBMOs)
- ) / numBMOsPerFragment;
+ double bmoMemoryRatio = bmoMemoryUsage / totalBMOsMemoryUsage;
+ double capMemoryRatio = 1;
+ if (totalNumBMOs > 1) {
+ capMemoryRatio =
ActiveSchemaDB()->getDefaults().getAsDouble(BMO_MEMORY_ESTIMATE_RATIO_CAP);
+ if (capMemoryRatio > 0 && capMemoryRatio <=1 && bmoMemoryRatio >
capMemoryRatio)
+ bmoMemoryRatio = capMemoryRatio;
--- End diff --
Yes. That's true. In that case, we need to do multiple pass or adjust the
totalBMOMemoryUsage for the unassigned BMO operator. Will attempt to do this
later
> Provide an improved memory quota assignment for big memory operators (BMO)
> --------------------------------------------------------------------------
>
> Key: TRAFODION-2733
> URL: https://issues.apache.org/jira/browse/TRAFODION-2733
> Project: Apache Trafodion
> Issue Type: Improvement
> Components: sql-cmp, sql-exe
> Affects Versions: 2.3-incubating
> Reporter: Selvaganesan Govindarajan
> Assignee: Selvaganesan Govindarajan
> Fix For: 2.3-incubating
>
>
> The big memory operators in Trafodion are HashJoin, HashGroupBy and Sort.
> Trafodion deploys multiple executor server processes (ESPs) to execute a
> query via its data flow architecture. Each ESPs can have an instance of this
> BMO operator. Currently, each instance of this operator can potentially have
> memory quota of 800 MB assigned to do its BMO operation. However, the memory
> allocation is usually limited by the memory pressure when this BMO attempts
> to allocate memory within the assigned quota. The assignment doesn't depend
> upon the estimation of memory needed by this operation.
> Improvement needed in BMO memory assignment are:
> 1. Limit the memory quota assignment for these BMO operations per node
> 2. Memory quota assigned taking into consideration estimated memory needed
> at every operator.
> 3. Ensure that the BMO gets the minimum memory needed at least to function
> smoothly
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
