>> Junichi Uekawa <[EMAIL PROTECTED]> writes: > Myrinet is coming very rapidly, and I think myrinet should be > considered. Well, that's my personal impression. (We have a cluster > with myrinet here).
Yup. Look at http://clusters.top500.org/ and play with the database queries. Some not so large clusters (32 nodes) are using either Myrinet or Myrinet 2000. The MPI implementation is either MPICH/GM or SCore. I dunno about GM's freeness. SCore is "almost" there (dunno about PM), it has some nasty attribution clauses[1] which push it to the non-free side[2]. Searching on the web I found some benchmarks which suggests MPICH/GM doesn't perform that well. The SCore implementation we have (Myrinet, 32bit, 33MHz) achieves 90 MB/s according to my own benchmarks. The newest one on Myrinet 2000/64bit/33MHz reaches 140 MB/s using the same benchmark. For some applications this makes sense. [1] I don't have a problem with attribution clauses. I do have a problem with those that say that you *must* mention you use someone's products on your papers. The other problem with SCore's license is that "you *must* report bugs to RWCP". [2] this was bad timing on my part: I just read SCore's license, I didn't know they distributed source, and just last week I was talking with Real World Computing's chief. -- Marcelo | "If you put butter and salt on it, it tastes like salty [EMAIL PROTECTED] | butter." | -- Popcorn comes to the Discworld | (Terry Pratchett, Moving Pictures)
