[
https://issues.apache.org/jira/browse/HIVE-9900?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14726178#comment-14726178
]
Siddharth Seth commented on HIVE-9900:
--------------------------------------
Posting comments here since ReviewBoard diff is currently broken.
{code}
80% may be too much, especially for smaller JVM heap sizes
{code}
80% may be a little high, given 15% is taken up by the cache. 70/75 would be a
more reliable value. This really depends on the overall JVM heap size though.
llapCluster.serviceStop(); - should be llapCluster.stop() (That takes care of
shutting down additional services which are part of the LlapDaemon composite
service)
{code}
+ llapDaemon = new LlapDaemon(conf, numExecutorsPerService,
execBytesPerService, llapIoEnabled,
+ ioIsDirect, ioBytesPerService, localDirs,
LlapConfiguration.LLAP_DAEMON_RPC_PORT_DEFAULT,
+ LlapConfiguration.LLAP_DAEMON_YARN_SHUFFLE_PORT_DEFAULT);
{code}
The ports should be left at 0, to set them up dynamically. Calling
getClusterSpecificConfiguration and merging it in with the main conf will
provide the randomly selected ports. Otherwise, multiple instances of the test
cannot run on the same node and causes problems with jenkins and other systems
which may be using the pre-defined default ports.
May want to set the queue size to a value below the default of 10.
Orthogonal to some extent: Is it possible to run the tests on MiniLlap with two
different configurations, one with a single executor and a small queue size.
This has worked very well in the past to detect deadlocks and scheduling issues.
Maybe this could be a command line parameter for the test - which can be run on
demand when required.
> LLAP: Integrate MiniLLAPCluster into tests
> ------------------------------------------
>
> Key: HIVE-9900
> URL: https://issues.apache.org/jira/browse/HIVE-9900
> Project: Hive
> Issue Type: Sub-task
> Reporter: Siddharth Seth
> Assignee: Prasanth Jayachandran
> Fix For: llap
>
> Attachments: HIVE-9900.1.patch, HIVE-MiniLlapCluster.patch
>
>
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
