pig-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Laukik Chitnis (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (PIG-1883) Pig's progress estimation should account for parallel job executions
Date Fri, 06 May 2011 00:37:03 GMT

    [ https://issues.apache.org/jira/browse/PIG-1883?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13029677#comment-13029677

Laukik Chitnis commented on PIG-1883:

> This doesn't lend itself well to automated testing. Any thoughts on how to test how the
new progress indicator does versus the existing one? Have you run any initial tests to measure

Thats correct; it is difficult to test in an automated fashion. One metric for defining the
performance of the progress estimator would be similar to what is used in the paratimer paper,
may be a RMS of the difference from the linear time (assuming "ideal" estimates are 0.0 to
1.0 from start time to finish time) 

I manually tested it with various pig scripts that generated different kinds of physical plans.
In most cases, I observed that the progress report was the same for both old and new methods.
One simple case where the new method does better is when a very small and a very large job
are executed in parallel. In this case, the old estimate shoots up to 50% very early, and
then moves slowly to 100%, whereas the new estimate grows more gradually from 0-100 as the
bigger job execution progresses. I haven't yet automated capturing these and analyzing the
metric yet.

> I don't understand the logic here. Why is it 0% done if ANY job is waiting, etc.? Some
of the jobs may be done and some partially done and some not even started.

The 0% is only for those set of jobs that are executing in parallel. For the set of jobs that
have finished execution in the previous rounds of parallel execution, their contribution to
the total estimate is 1/#rounds per round of execution i.e. per JobControl object (so, #rounds
is the length of the critical path along the operator plan tree)

> This code shouldn't be in OperatorPlan. We want to keep that as clean as possible. Instead
you should build a new Walker type that can do this calculation.

Ah, ok; Will do that.

> You have tabs here and some other spots. Please make sure you use 4 spaces rather than

I need to change my editor's auto-indentation formatting :)

> Why is a separate method needed here? When users turn on the new progress indicator I
assume they don't get the old one too. Given that the interfaces are the same it seems one
method should suffice here.

Initially, I put in a separate method assuming that users could have listeners for either
of them. For example, we could use these separate listeners for the performance comparison
between the old and new methods. Later on, however, when I added the command-line option to
choose, I made the old and new methods as an either-or choice. Perhaps I should make it possible
to have both indicators turned on at the same time?

> It looks like this comment got attached to the run method. Also, the method has only
one parameter, but two are listed in the comment.

I will fix this.

> Pig's progress estimation should account for parallel job executions
> --------------------------------------------------------------------
>                 Key: PIG-1883
>                 URL: https://issues.apache.org/jira/browse/PIG-1883
>             Project: Pig
>          Issue Type: Improvement
>            Reporter: Laukik Chitnis
>            Assignee: Laukik Chitnis
>         Attachments: PIG-1883-2.patch
> Currently, Pig's progress estimation is based on the percentage of jobs completed out
of the total number of MR jobs. However, since the MR operators are arranged in a DAG (and
hence more than 1 job might be submitted for execution in parallel), the progress estimation
can be improved by considering the number of jobs in the critical path, instead of just the
total number of jobs.

This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira

View raw message