[
https://issues.apache.org/jira/browse/GEODE-6084?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Dan Smith updated GEODE-6084:
-----------------------------
Description:
We currently have a test PartitionedPutBenchmark class, but this has all
threads updating the same integer key. We need to turn this into a real,
reasonable benchmark of puts for Geode.
This is the configuration we want to go for:
*Workload time*
As long as we need to get reasonable stddev numbers, targeting around 5
minutes or less.
*Topology*
Client/server, with single hop enabled and no SSL
*Data*
* PDX object with around a 1K serialized size.
* Uniformly and randomly accessed/updated
* as many entries as we can reasonably populate for a 5 minute test. Long or
small string keys.
* The data should be pre-populated before the workload starts.
*Scale*
* 4 servers, enough clients and threads to be utilizing 50-100% of the
critical resource (CPU, network?)
*Region Config*
Create a test for both
1. Partitioned Region with redundancy 1
2. Replicated Regions
*Geode Tuning*
We should use Geode and JVM tuning parameters tuned for performance. This
includes
*Geode properties*
{noformat}
eviction-heap-percentage=65
critical-heap-percentage=90
archive-disk-space-limit=150
archive-file-size-limit=10
conserve-sockets=false
enable-time-statistics=true
jmx-manager=true
jmx-manager-port=1099
jmx-manager-start=true
jmx-manager-update-rate=2000
locator-wait-time=120
log-disk-space-limit=100
log-file-size-limit=10
log-level=config
remove-unresponsive-client=true
statistic-archive-file=XXXX
statistic-sample-rate=1000
statistic-sampling-enabled=true
log-file=XXXX
distributed-system-id=0
{noformat}
*JVM flags*
{noformat}
-XX:CMSInitiatingOccupancyFraction=60
-XX:+PrintGCDetails
-XX:+PrintGCTimeStamps
-XX:+PrintGCDateStamps
-XX:+PrintGCApplicationStoppedTime
-XX:+PrintGCApplicationConcurrentTime
-XX:+UseGCLogFileRotation
-XX:NumberOfGCLogFiles=20
-XX:GCLogFileSize=1M
-XX:+UnlockDiagnosticVMOptions
-XX:ParGCCardsPerStrideChunk=32768
'-XX:OnOutOfMemoryError='\''kill -9 %p'\'''
-XX:+UseNUMA
-XX:+UseConcMarkSweepGC
-XX:+UseCMSInitiatingOccupancyOnly
-XX:+CMSClassUnloadingEnabled
-XX:+DisableExplicitGC
-XX:+ScavengeBeforeFullGC
-XX:+CMSScavengeBeforeRemark
-server
-Djava.awt.headless=true
-Dsun.rmi.dgc.server.gcInterval=9223372036854775806
-Dgemfire.OSProcess.ENABLE_OUTPUT_REDIRECTION=true
-Dgemfire.launcher.registerSignalHandlers=true
{noformat}
was:
We currently have a test PartitionedPutBenchmark class, but this has all
threads updating the same integer key. We need to turn this into a real,
reasonable benchmark of puts for Geode.
This is the configuration we want to go for:
*Workload time*
As long as we need to get reasonable stddev numbers, targeting around 5
minutes or less.
*Topology*
Client/server, with single hop enabled and no SSL
*Data*
* PDX object with around a 1K serialized size.
* Uniformly and randomly accessed/updated
* as many entries as we can reasonably populate for a 5 minute test. Long or
small string keys.
* The data should be pre-populated before the workload starts.
*Scale*
* 4 servers, enough clients and threads to be utilizing 50-100% of the
critical resource (CPU, network?)
*Region Config*
Create a test for both
1. Partitioned Region with redundancy 1
2. Replicated Regions
*Geode Tuning*
We should use Geode and JVM tuning parameters tuned for performance. This
includes
*Geode properties*
eviction-heap-percentage=65
critical-heap-percentage=90
archive-disk-space-limit=150
archive-file-size-limit=10
conserve-sockets=false
enable-time-statistics=true
jmx-manager=true
jmx-manager-port=1099
jmx-manager-start=true
jmx-manager-update-rate=2000
locator-wait-time=120
log-disk-space-limit=100
log-file-size-limit=10
log-level=config
remove-unresponsive-client=true
statistic-archive-file=XXXX
statistic-sample-rate=1000
statistic-sampling-enabled=true
log-file=XXXX
distributed-system-id=0
*JVM flags*
{noformat}
-XX:CMSInitiatingOccupancyFraction=60
-XX:+PrintGCDetails
-XX:+PrintGCTimeStamps
-XX:+PrintGCDateStamps
-XX:+PrintGCApplicationStoppedTime
-XX:+PrintGCApplicationConcurrentTime
-XX:+UseGCLogFileRotation
-XX:NumberOfGCLogFiles=20
-XX:GCLogFileSize=1M
-XX:+UnlockDiagnosticVMOptions
-XX:ParGCCardsPerStrideChunk=32768
'-XX:OnOutOfMemoryError='\''kill -9 %p'\'''
-XX:+UseNUMA
-XX:+UseConcMarkSweepGC
-XX:+UseCMSInitiatingOccupancyOnly
-XX:+CMSClassUnloadingEnabled
-XX:+DisableExplicitGC
-XX:+ScavengeBeforeFullGC
-XX:+CMSScavengeBeforeRemark
-server
-Djava.awt.headless=true
-Dsun.rmi.dgc.server.gcInterval=9223372036854775806
-Dgemfire.OSProcess.ENABLE_OUTPUT_REDIRECTION=true
-Dgemfire.launcher.registerSignalHandlers=true
{noformat}
> Create real put benchmark for PDX values
> ----------------------------------------
>
> Key: GEODE-6084
> URL: https://issues.apache.org/jira/browse/GEODE-6084
> Project: Geode
> Issue Type: New Feature
> Components: benchmarks
> Reporter: Dan Smith
> Priority: Major
> Labels: 3_points
>
> We currently have a test PartitionedPutBenchmark class, but this has all
> threads updating the same integer key. We need to turn this into a real,
> reasonable benchmark of puts for Geode.
> This is the configuration we want to go for:
> *Workload time*
> As long as we need to get reasonable stddev numbers, targeting around 5
> minutes or less.
> *Topology*
> Client/server, with single hop enabled and no SSL
> *Data*
> * PDX object with around a 1K serialized size.
> * Uniformly and randomly accessed/updated
> * as many entries as we can reasonably populate for a 5 minute test. Long or
> small string keys.
> * The data should be pre-populated before the workload starts.
> *Scale*
> * 4 servers, enough clients and threads to be utilizing 50-100% of the
> critical resource (CPU, network?)
> *Region Config*
> Create a test for both
> 1. Partitioned Region with redundancy 1
> 2. Replicated Regions
> *Geode Tuning*
> We should use Geode and JVM tuning parameters tuned for performance. This
> includes
>
> *Geode properties*
> {noformat}
> eviction-heap-percentage=65
> critical-heap-percentage=90
> archive-disk-space-limit=150
> archive-file-size-limit=10
> conserve-sockets=false
> enable-time-statistics=true
> jmx-manager=true
> jmx-manager-port=1099
> jmx-manager-start=true
> jmx-manager-update-rate=2000
> locator-wait-time=120
> log-disk-space-limit=100
> log-file-size-limit=10
> log-level=config
> remove-unresponsive-client=true
> statistic-archive-file=XXXX
> statistic-sample-rate=1000
> statistic-sampling-enabled=true
> log-file=XXXX
> distributed-system-id=0
> {noformat}
> *JVM flags*
> {noformat}
> -XX:CMSInitiatingOccupancyFraction=60
> -XX:+PrintGCDetails
> -XX:+PrintGCTimeStamps
> -XX:+PrintGCDateStamps
> -XX:+PrintGCApplicationStoppedTime
> -XX:+PrintGCApplicationConcurrentTime
> -XX:+UseGCLogFileRotation
> -XX:NumberOfGCLogFiles=20
> -XX:GCLogFileSize=1M
> -XX:+UnlockDiagnosticVMOptions
> -XX:ParGCCardsPerStrideChunk=32768
> '-XX:OnOutOfMemoryError='\''kill -9 %p'\'''
> -XX:+UseNUMA
> -XX:+UseConcMarkSweepGC
> -XX:+UseCMSInitiatingOccupancyOnly
> -XX:+CMSClassUnloadingEnabled
> -XX:+DisableExplicitGC
> -XX:+ScavengeBeforeFullGC
> -XX:+CMSScavengeBeforeRemark
> -server
> -Djava.awt.headless=true
> -Dsun.rmi.dgc.server.gcInterval=9223372036854775806
> -Dgemfire.OSProcess.ENABLE_OUTPUT_REDIRECTION=true
> -Dgemfire.launcher.registerSignalHandlers=true
> {noformat}
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
