[
https://issues.apache.org/jira/browse/YARN-415?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Eric Payne updated YARN-415:
----------------------------
Attachment: YARN-415.201407232237.txt
[~leftnoteasy], Thank you for your reply.
I have implemented the following changes with the current patch.
{quote}
1. Revert changes of SchedulerAppReport, we already have changed
ApplicationResourceUsageReport, and memory utilization should be a part of
resource usage report.
{quote}
Changes to SchedulerAppReport have been reverted.
{quote}
2. Remove getMemory(VCore)Seconds from RMAppAttempt, modify
RMAppAttemptMetrics#getFinishedMemory(VCore)Seconds to return completed+running
resource utilization.
{quote}
I have removed getters and setters from RMAppAttempt and added
RMAppAttemptMetrics#getResourceUtilization, which returns a single
ResourceUtilization instance that contains both memorySeconds and vcoreSeconds
for the appAttempt. These include both finished and running statistics IF the
appAttempt is ALSO the current attempt. If not, it only includes the finished
statistics.
{quote}
3. put
{code}
._("Resources:",
String.format("%d MB-seconds, %d vcore-seconds",
app.getMemorySeconds(), app.getVcoreSeconds()))
{code}
from "Application Overview" to "Application Metrics", and rename it to
"Resource Seconds". It should be considered as a part of application metrics
instead of overview.
{quote}
Changes completed.
{quote}
4. Change finishedMemory/VCoreSeconds to AtomicLong in RMAppAttemptMetrics to
make it can be efficiently accessed by multi-thread.
{quote}
Changes completed.
{quote}
5. I think it's better to add a new method in SchedulerApplicationAttempt like
getMemoryUtilization, which will only return memory/cpu seconds. We do this to
prevent locking scheduling thread when showing application metrics on web UI.
getMemoryUtilization will be used by
RMAppAttemptMetrics#getFinishedMemory(VCore)Seconds to return completed+running
resource utilization. And used by
SchedulerApplicationAttempt#getResourceUsageReport as well.
The MemoryUtilization class may contain two fields:
runningContainerMemory(VCore)Seconds.
{quote}
Added ResourceUtilization (instead of MemoryUtilization), but did not make the
other changes as per comment:
https://issues.apache.org/jira/browse/YARN-415?focusedCommentId=14071181&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-14071181
{quote}
6. Since compute running container resource utilization is not O(1), we need
scan all containers under an application. I think it's better to cache a
previous compute result, and it will be recomputed after several seconds (maybe
1-3 seconds should be enough) elapsed.
{quote}
I added chached values in SchedulerApplicationAttempt for memorySeconds and
vcoreSeconds that are updated when 1) a request is received to calculate these
metrics, AND 2) it has been more than 3 seconds since the last request.
One thing I did notice when these values are cached is that there is a race
where containers can get counted twice:
- RMAppAttemptMetrics#getResourceUtilization sends a request to calculate
running containers, and container X is almost finished.
RMAppAttemptMetrics#getResourceUtilization adds the finished values to the
running values and returns ResourceUtilization.
- Container X completes and its memorySeconds and vcoreSeconds are added to the
finished values for appAttempt.
- RMAppAttemptMetrics#getResourceUtilization makes another request before the 3
second interval, and the cached values are added to the finished values for
appAttempt.
Since both the cached values and the finished values contain metrics for
Container X, those are double counted until 3 seconds elapses and the next
RMAppAttemptMetrics#getResourceUtilization request is made.
{quote}
And you can modify SchedulerApplicationAttempt#liveContainers to be a
ConcurrentHashMap. With #6, get memory utilization to show metrics on web UI
will not lock scheduling thread at all.
{quote}
I am a little reluctant to modify the type of
SchedulerApplicationAttempt#liveContainers as part of this JIRA. That seems
like something that could be done separately.
> Capture memory utilization at the app-level for chargeback
> ----------------------------------------------------------
>
> Key: YARN-415
> URL: https://issues.apache.org/jira/browse/YARN-415
> Project: Hadoop YARN
> Issue Type: New Feature
> Components: resourcemanager
> Affects Versions: 0.23.6
> Reporter: Kendall Thrapp
> Assignee: Andrey Klochkov
> Attachments: YARN-415--n10.patch, YARN-415--n2.patch,
> YARN-415--n3.patch, YARN-415--n4.patch, YARN-415--n5.patch,
> YARN-415--n6.patch, YARN-415--n7.patch, YARN-415--n8.patch,
> YARN-415--n9.patch, YARN-415.201405311749.txt, YARN-415.201406031616.txt,
> YARN-415.201406262136.txt, YARN-415.201407042037.txt,
> YARN-415.201407071542.txt, YARN-415.201407171553.txt,
> YARN-415.201407172144.txt, YARN-415.201407232237.txt, YARN-415.patch
>
>
> For the purpose of chargeback, I'd like to be able to compute the cost of an
> application in terms of cluster resource usage. To start out, I'd like to
> get the memory utilization of an application. The unit should be MB-seconds
> or something similar and, from a chargeback perspective, the memory amount
> should be the memory reserved for the application, as even if the app didn't
> use all that memory, no one else was able to use it.
> (reserved ram for container 1 * lifetime of container 1) + (reserved ram for
> container 2 * lifetime of container 2) + ... + (reserved ram for container n
> * lifetime of container n)
> It'd be nice to have this at the app level instead of the job level because:
> 1. We'd still be able to get memory usage for jobs that crashed (and wouldn't
> appear on the job history server).
> 2. We'd be able to get memory usage for future non-MR jobs (e.g. Storm).
> This new metric should be available both through the RM UI and RM Web
> Services REST API.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
