hadoop-yarn-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Jason Lowe (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (YARN-7136) Additional Performance Improvement for Resource Profile Feature
Date Fri, 01 Sep 2017 15:08:00 GMT

    [ https://issues.apache.org/jira/browse/YARN-7136?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16150688#comment-16150688
] 

Jason Lowe commented on YARN-7136:
----------------------------------

Thanks for updating the patch!

The findbugs exclusion entry should just change the filename for BaseResource to Resource
rather than add a new entry, since BaseResource no longer has the method.

ResourcePBImpl's getResources method is redundant since it simply delegates to the super method.

Resource's compareTo is now treating all the resources equally which is nice, but that means
for memory and vcores it will be doing string compares on the names, hash set lookups on the
units, and string comparisons on the units before it ever gets to simply comparing the numbers
which is all we really care about for these resources.  As I've mentioned in offline conversation,
I don't understand why we would ever want to do any unit comparisons in the scheduler allocation
code.  It may make sense from a performance standpoint to have something like a NormalizedResource
that has pre-normalized all the ResourceInformation entries to a base precision unit for that
resource so the scheduler can avoid doing any resource name comparisons (the resource index
match should be an equivalent test) and any unit conversions.  That would avoid all string
comparisons and hash set lookups for resource comparison in the scheduler main-path code,
reducing it to simply a loop over an array of numbers we can directly compare without further
computation.  The ApplicationMasterService can marshal the app's requests into the normalized
resources before feeding the requests to the scheduler so the scheduler doesn't have to ever
worry about units.  Worth a followup JIRA?

Similar performance hit for the Resources equals method since it now needs to check names
and units before checking values on memory and vcores which it did not do before.

For the toString change this will start to show units for memory ("Mi") which was not done
before.  Does memory need to have no units for backwards-compatibility as it is for vcores?
 Otherwise it seems like as soon as a custom resources is added memory will suddenly go from
no units to "Mi" units in toString as we switch from BaseResource.

In many places in DominantResourceCalculator:
{code}
    for (int i = 0; i < ResourceUtils.getNumberOfKnownResourceTypes(); i++) {
{code}
getNumberOfKnownResourceTypes has to touch a volatile variable each time, so we should cache
this in a local rather than hit it every time in the loop.  Speaking of hitting this volatile
variable, what happens if this value suddenly changes to be larger than the array of resources
passed in?  In other words, it seems broken that we're looking at a global variable to determine
how many entries in the array argument to look at as they could be inconsistent.  Thinking
that if an admin suddenly refreshes the resource types to add one the code could end up indexing
off the end of an array argument that was just passed to it before the refresh occurred.

DominantResourceCalculator#calculateSharesForMandatoryResources says it is custom-built to
only check two resources and assumes clusterRes only has two entries yet is calling ResourceUtils.getNumberOfKnownResourceTypes()
and iterating that many times through clusterRes?

Same hitting-a-volatile-in-the-loop issue for Resources.

> Additional Performance Improvement for Resource Profile Feature
> ---------------------------------------------------------------
>
>                 Key: YARN-7136
>                 URL: https://issues.apache.org/jira/browse/YARN-7136
>             Project: Hadoop YARN
>          Issue Type: Sub-task
>          Components: nodemanager, resourcemanager
>            Reporter: Wangda Tan
>            Assignee: Wangda Tan
>            Priority: Critical
>         Attachments: YARN-7136.001.patch, YARN-7136.YARN-3926.001.patch, YARN-7136.YARN-3926.002.patch,
YARN-7136.YARN-3926.003.patch, YARN-7136.YARN-3926.004.patch, YARN-7136.YARN-3926.005.patch
>
>
> This JIRA is plan to add following misc perf improvements:
> 1) Use final int in Resources/ResourceCalculator to cache #known-resource-types. (Significant
improvement).
> 2) Catch Java's ArrayOutOfBound Exception instead of checking array.length every time.
(Significant improvement).
> 3) Avoid setUnit validation (which is a HashSet lookup) when initialize default Memory/VCores
ResourceInformation (Significant improvement).
> 4) Avoid unnecessary loop array in Resource#toString/hashCode. (Some improvement).
> 5) Removed readOnlyResources in BaseResource. (Minor improvement).
> 6) Removed enum: MandatoryResources, use final integer instead. (Minor improvement).



--
This message was sent by Atlassian JIRA
(v6.4.14#64029)

---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscribe@hadoop.apache.org
For additional commands, e-mail: yarn-issues-help@hadoop.apache.org


Mime
View raw message