You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@singa.apache.org by "wangwei (JIRA)" <ji...@apache.org> on 2016/01/13 07:20:39 UTC
[jira] [Updated] (SINGA-135) Improve hybrid partitioning and
benchmark performance
[ https://issues.apache.org/jira/browse/SINGA-135?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
wangwei updated SINGA-135:
--------------------------
Description:
SINGA now supports partitioning on batch dimension and feature dimension (SINGA-113). Given more than two workers per group, every layer is partitioned. It is possible that some layers are not suitable for partitioning, e.g., the SoftmaxLossLayer in Alexnet. To handle that case, a special value should be configured for that layer's partition dimension field, e.g., -1.
With all partitioning dimensions implemented, we can bench mark the performance of different partitioning schemes. The benchmark task could be in a new ticket or in this ticket.
was:
SINGA now supports partitioning on batch dimension and feature dimension. Given more than two workers per group, every layer is partitioned. It is possible that some layers are not suitable for partitioning, e.g., the SoftmaxLossLayer in Alexnet. To handle that case, a special value should be configured for that layer's partition dimension field, e.g., -1.
With all partitioning dimensions implemented, we can bench mark the performance of different partitioning schemes. The benchmark task could be in a new ticket or in this ticket.
> Improve hybrid partitioning and benchmark performance
> -----------------------------------------------------
>
> Key: SINGA-135
> URL: https://issues.apache.org/jira/browse/SINGA-135
> Project: Singa
> Issue Type: Improvement
> Reporter: wangwei
> Assignee: Sheng Wang
>
> SINGA now supports partitioning on batch dimension and feature dimension (SINGA-113). Given more than two workers per group, every layer is partitioned. It is possible that some layers are not suitable for partitioning, e.g., the SoftmaxLossLayer in Alexnet. To handle that case, a special value should be configured for that layer's partition dimension field, e.g., -1.
> With all partitioning dimensions implemented, we can bench mark the performance of different partitioning schemes. The benchmark task could be in a new ticket or in this ticket.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)