You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@systemml.apache.org by "Matthias Boehm (JIRA)" <ji...@apache.org> on 2018/02/03 08:41:00 UTC
[jira] [Created] (SYSTEMML-2120) Improve codegen optimizer (pruning
effectiveness)
Matthias Boehm created SYSTEMML-2120:
----------------------------------------
Summary: Improve codegen optimizer (pruning effectiveness)
Key: SYSTEMML-2120
URL: https://issues.apache.org/jira/browse/SYSTEMML-2120
Project: SystemML
Issue Type: Sub-task
Reporter: Matthias Boehm
The codegen optimizer applies various techniques to prune the search space, including the partitioning into independent problems, the restriction to interesting points, as well as cost-based and structural pruning. These techniques work very well for most use case rendering the codegen overhead negligible. However, there are still exceptions where the additional compilation overhead leads to a slowdown. One of these examples is lenet, the below codegen statistics are obtained from 1000 iterations over mnist60k.
{code}
HOP DAGs recompiled (PRED, SB): 0/1154.
HOP DAGs recompile time: 237.579 sec.
Codegen compile (DAG,CP,JC): 1176/17450/36.
Codegen enum (ALLt/p,EVALt/p): 1064354993445/16387318/11342971/11331811.
Codegen compile times (DAG,JC): 234.174/0.835 sec.
Codegen plan cache hits: 17408/17444.
{code}
Pruning reduced the number of plans from > 1 trillion to 11 million, which is still too high causing an overhead of 234s for a total execution time of 425s.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)