[Beowulf] RDMA NICs and future beowulfs
diep at xs4all.nl
Wed Apr 27 03:34:04 PDT 2005
Nothing new in programming.
In general when using MPI commands even doctors in computer science manage
to slow down their software factor 40 first, in order to get better scaling.
In computerchess a very normal phenomena.
Others manage to get a worse exponent even, like Deep Blue did,
using wrong algorithms.
At 10:01 AM 4/26/2005 -0500, Ben Mayer wrote:
>Just a quick comment about CELL and gcc and vectorizing.
>I looked over the specs of CELL and it looks very much like a hybrid
>vector processor. gcc now has support for vectorizing (some) loops.
>I spend a good deal of time writing code for Cray X1Es (vector
>supercomputer), and there are a lot of things that can trip up a
>compiler when one is trying to get it to vectorize code. I actually
>just did a small study of how well students in a parallel computing
>class write parallel codes on X1 with MPI and UPC. One of the things
>that stood out is that they tended to do odd things in their loops
>that inhibit code from vectorizing.
>Writing code which vectorizes well is difficult. The reason I mention
>this is that, I believe, we will see that someone *has to* produce
>libraries (or entire engines) that take advantage of the Cells unique
>layout to get decent performance. If there are major parts of the code
>which can not run in parallel (Vectorization is a type of parallel
>execution) Cell is just a crummy PPC chip.
>Beowulf mailing list, Beowulf at beowulf.org
>To change your subscription (digest mode or unsubscribe) visit
More information about the Beowulf