Presentation from a summit on software and algorithms for the Cell Processor

[Colourless]

Ran it on a SPU (fixed some bugs first) it is actually slower that the original. It's longer by 4 instructions. And introduces the following instructions to the pipeline 2x ceqi (compare equal) and 2x selb (select bits).

Heh. Good thing I didn't say it should be faster

 

[patsu]

Ok, this is the missing workshop slides in the OP (although everything else is gone by now).

Talks about using Cell to speed up Breadth-First-Search:
http://storeandserve.com/download/673577/cell-workshop-petrini.pdf.html?fmw=WQ==

Problem Description:

Breadth-First Graph (BFS) Exploration
-- Random memory access pattern (when exploring random graphs, which is our case)
-- Fine granularity (tens, hundreds of bytes)
-- Little opportunities to increase the communication granularity
-- Almost no computation
-- Reading/updating bitmap and data transfers
---- Difficult problem for conventional processors
---- Can’t get any worse in terms of irregular access pattern and communication granularity

Thanks to Fabrizio Petrini for sharing.

 

[Bohdy]

The BFS results are extremely impressive. Cell is really approaching the supercomputer territory in that one.

 

[patsu]

Yes, this is supposed to be a problem unsuitable for the Cell architecture.

The author says more coming up !

 

[Shifty Geezer]

The BFS results are extremely impressive. Cell is really approaching the supercomputer territory in that one.

Gobsmacking if they're legitimate! The performance is way beyond other single chip computers, rivalling 128 and 256 CPU Blue Gene supercomputers, and set to give that MTA-2 a run for it's money too. It'd be nice to compare performance to a standard implementation to show how targetting the Cell architecture improves things, and another look at scaling the process over multiple Cells. ie. When you lose the locality of the ring bus, has does scaling CPU numbers affect scaling of results?

 

[Carl B]

Shifty, yeah, I'd be interested in seeing how the Cell scales as well; it's 'per chip' performance is awesome though. Great find Patsu, thanks for the link! By the way, this may be a topic for another thread, but do we know if along with the DP enhancements in the pipe, Cell revisions with a larger LS are in the cards?

 

[DennyUA]

The forum has liked...

The good forum, has found answers to many questions.
Thanks)