[jira] [Commented] (FLINK-1195) Improvement of benchmarking infrastructure

["Till Rohrmann (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/FLINK-1195?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14204922#comment-14204922 ] 

Till Rohrmann commented on FLINK-1195:
--------------------------------------

I looked briefly over Peel and I agree with you. It looks like a potential solution for our problem. At the moment we are running the following benchmarks: 

Grep: https://github.com/project-flink/flink-perf/tree/master/flink-jobs/src/main/java/com/github/projectflink/grep
Pagerank: https://github.com/project-flink/flink-perf/tree/master/flink-jobs/src/main/java/com/github/projectflink/pagerank
GroupReduce: https://github.com/project-flink/flink-perf/blob/master/flink-jobs/src/main/scala/com/github/projectflink/GroupReduceBenchmark.scala
ALS: https://github.com/project-flink/flink-perf/tree/master/flink-jobs/src/main/scala/com/github/projectflink/als

And for comparison we run

GroupReduce: https://github.com/project-flink/flink-perf/tree/master/spark-jobs/src/main/scala/com/github/projectflink/spark/GroupReduceBenchmark.scala
ALS: https://github.com/project-flink/flink-perf/tree/master/spark-jobs/src/main/scala/com/github/projectflink/spark/als

on Spark as well. It would be really great if we could automate the benchmarking process.

> Improvement of benchmarking infrastructure
> ------------------------------------------
>
>                 Key: FLINK-1195
>                 URL: https://issues.apache.org/jira/browse/FLINK-1195
>             Project: Flink
>          Issue Type: Wish
>            Reporter: Till Rohrmann
>            Assignee: Alexander Alexandrov
>
> I noticed while running my ALS benchmarks that we still have some potential to improve our benchmarking infrastructure. The current state is that we execute the benchmark jobs by writing a script with a single set of parameters. The runtime is then manually retrieved from the web interface of Flink and Spark, respectively.
> I think we need the following extensions:
> * Automatic runtime retrieval and storage in a file
> * Repeated execution of jobs to gather some "advanced" statistics such as mean and standard deviation of the runtimes
> * Support for value sets for the individual parameters
> The automatic runtime retrieval would allow us to execute several benchmarks consecutively without having to lookup the runtimes in the logs or in the web interface, which btw only stores the runtimes of the last 5 jobs.
> What I mean with value sets is that would be nice to specify a set of parameter values for which the benchmark is run without having to write for every single parameter combination a benchmark script. I believe that this feature would become very handy when we want to look at the runtime behaviour of Flink for different input sizes or degrees of parallelism, for example. To illustrate what I mean:
> {code}
> INPUTSIZE = 1000, 2000, 4000, 8000
> DOP = 1, 2, 4, 8
> OUTPUT=benchmarkResults
> repetitions=10
> command=benchmark.jar -p $DOP $INPUTSIZE 
> {code} 
> Something like that would execute the benchmark job with (DOP=1, INPUTSIZE=1000), (DOP=2, INPUTSIZE=2000),.... 10 times each, calculate for each parameter combination runtime statistics and store the results in the file benchmarkResults.
> I believe that spending some effort now will pay off in the long run because we will benchmark Flink continuously. What do you guys think?



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)