singa-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "ASF subversion and git services (JIRA)" <>
Subject [jira] [Commented] (SINGA-24) Implement Downpour training framework
Date Fri, 26 Jun 2015 12:51:05 GMT


ASF subversion and git services commented on SINGA-24:

Commit 14ce5d9aee5d9c4a2cd6c7bc3a64ae3df4f5902f in incubator-singa's branch refs/heads/master
from wang wei
[;h=14ce5d9 ]

SINGA-24 Implement Downpour training framework

Downpour training framwork has multiple worker groups and single server groups.
Note: Param slices of servers would share memory space with local workers.
If the local worker is not from group 0 who does the put requests, but it has Param slices
in local servers,
then it has to tell local servers in the Get requests the pointers for the shared slices's
memory space.

Tested with worker_server_separate= true and false, server/worker group with one and more

> Implement Downpour training framework
> -------------------------------------
>                 Key: SINGA-24
>                 URL:
>             Project: Singa
>          Issue Type: New Feature
>            Reporter: wangwei
> The downpour training framework is discussed in Google Brain.
> Multiple worker groups compute gradients of parameters asynchronously. A single server
group which many have multiple workers conduct update for parameters. The servers and workers
may resident in the same process or different processes depending on cluster configuration
parameter(i.e., worker_server_separate).

This message was sent by Atlassian JIRA

View raw message