Actually, data distribution method is key. Cannon's, FHO, DNS algorithms are all based on block-cyclic data distribution or can be used for limited cases such as processor num equals power of 2 or square number.
See also SUMMA or PUMMA: http://www.cs.utexas.edu/ftp/techreports/tr95-13.pdf On Mon, Mar 26, 2012 at 1:58 PM, Thomas Jungblut (Commented) (JIRA) <[email protected]> wrote: > > [ > https://issues.apache.org/jira/browse/HAMA-221?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13238102#comment-13238102 > ] > > Thomas Jungblut commented on HAMA-221: > -------------------------------------- > > Hmm you probably need Java7 for that. But I guess you could remove my vector > class with that of mahout-math. I have quite the same interface. > > And I think it is not very scalable. So a DNS would be better. > Heres a paper about how it could look like: > http://cloudscale.com/images/bsp/matmultbsp.pdf > >> A BSP matrix multiplication implementation. >> ------------------------------------------- >> >> Key: HAMA-221 >> URL: https://issues.apache.org/jira/browse/HAMA-221 >> Project: Hama >> Issue Type: Improvement >> Components: bsp, matrix >> Affects Versions: 0.3.0 >> Reporter: Edward J. Yoon >> Assignee: Thomas Jungblut >> >> Using the BSP model, we could improve the performance of mat-mat >> multiplication. Let's evaluation it out on this issue. > > -- > This message is automatically generated by JIRA. > If you think it was sent incorrectly, please contact your JIRA > administrators: > https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa > For more information on JIRA, see: http://www.atlassian.com/software/jira > > -- Best Regards, Edward J. Yoon @eddieyoon
