Maxim Khutornenko created AURORA-1156:
-----------------------------------------
Summary: Preemptor perf improvements
Key: AURORA-1156
URL: https://issues.apache.org/jira/browse/AURORA-1156
Project: Aurora
Issue Type: Task
Components: Scheduler
Reporter: Maxim Khutornenko
There are two minor algorithm changes that should result in overall preemptor
perf improvements:
1. There seems to be a redundant scheduling attempt here: \[1\]. The preemptor
loop is triggered within the same storage write lock right after a failed
{{OfferManager.launchFirst()}} run that has already attempted to match all
available offers. There is no reason to attempt another match given the same
system conditions, especially considering that a positive scheduling result
from this loop is discarded anyway \[2\].
2. Instead of running an expensive scheduling loop \[3\] for every task victim,
it's possible to "size up" victim resources and call scheduling filter only
when rough resource estimates are sufficient. Something like:
{noformat}
if (totalResource.greaterOrEqualThan(ResourceSlot.from(pendingTask))) {
// call schedulingFilter
}
{noformat}
\[1\] -
https://github.com/apache/incubator-aurora/blob/master/src/main/java/org/apache/aurora/scheduler/async/preemptor/PreemptorImpl.java#L220-L237
\[2\] -
https://github.com/apache/incubator-aurora/blob/master/src/main/java/org/apache/aurora/scheduler/async/preemptor/PreemptorImpl.java#L338
\[3\] -
https://github.com/apache/incubator-aurora/blob/master/src/main/java/org/apache/aurora/scheduler/async/preemptor/PreemptorImpl.java#L263-L265
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
