[
https://issues.apache.org/jira/browse/SYSTEMML-845?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15437647#comment-15437647
]
Mike Dusenberry commented on SYSTEMML-845:
------------------------------------------
So here's the basic summary again:
* 2 scripts, one without DML functions, one with DML functions.
* Same performance in forced singlenode mode.
* Slower performance in script with DML functions running in regular hybrid
modes due to extra MR or Spark job creation.
> Compare Performance of LeNet Scripts With & Without Using SystemML-NN
> ---------------------------------------------------------------------
>
> Key: SYSTEMML-845
> URL: https://issues.apache.org/jira/browse/SYSTEMML-845
> Project: SystemML
> Issue Type: Improvement
> Components: Algorithms, Compiler
> Affects Versions: SystemML 0.11
> Reporter: Mike Dusenberry
> Assignee: Mike Dusenberry
> Attachments: convert.dml,
> lenet-train-spark-explain-recompile-hops.log, lenet-train-spark-explain.log,
> log08.03.16-1470268602.txt,
> mnist_lenet-train-spark-explain-recompile-hops.log,
> mnist_lenet-train-spark-explain.log, perf.sh, run.sh
>
>
> This JIRA issue tracks the comparison of the performance of the LeNet scripts
> with & without using SystemML-NN. The goal is that they should have equal
> performance in terms of both accuracy and time. Any difference will be
> indicate areas of engine improvement.
> Scripts:
> * [mnist_lenet-train.dml |
> https://github.com/apache/incubator-systemml/blob/master/scripts/staging/SystemML-NN/examples/mnist_lenet-train.dml]
> - LeNet script that *does* use the SystemML-NN library.
> * [lenet-train.dml |
> https://github.com/apache/incubator-systemml/blob/master/scripts/staging/lenet-train.dml]
> - LeNet script that *does not* use the SystemML-NN library.
> *Current Status - Forced Singlenode:*
> Equal performance when running the scripts in standalone mode with the
> {{-exec singlenode}} flag, 20GB of memory, and using data inputs in the
> SystemML binary format -- see {{run.sh}} and {{perf.sh}} for information.
> Results:
> - Run #1:
> || Script | Time (s) | Accuracy ||
> | mnist_lenet-train.dml | 2987.400704441 | 99.32% |
> | lenet-train.dml | 2816.369435579 | 99.28% |
> - Run #2:
> || Script | Time (s) | Accuracy ||
> | mnist_lenet-train.dml | 2847.790531812 | 99.16% |
> | lenet-train.dml | 2950.520494210 | 99.18% |
> So, same accuracy, and same runtime in singlenode mode!
> To fully reproduce, I basically created a directory, placed the two attached
> bash scripts in it, grabbed a copy of the NN library and placed it into the
> directory, ran the examples/get_mnist_data.sh script from the library to get
> the data (placed into examples/data), then used the attached convert.dml to
> create binary copies of the data for both scripts, then ran run.sh. Also, I
> copied examples/data to the base directory as well. Adjust the {{EXEC}} and
> related variables in {{perf.sh}} to switch between standalone, Spark, memory
> sizes, explain, stats, etc.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
