[Beowulf] help on building Beowulf
bill at cse.ucdavis.edu
Wed Nov 21 00:22:53 PST 2007
I ran a 32 node/64 cpu and a 16 node 32 cpu (although I tinkered with
hyperthreading for twice that) for 3 years or so.
Mryinet was stable, did occasionally have problems with job clean up, but I've
seen similar with infinipath, infiniband, and even ethernet. We had a few
support calls and got our issues resolved quickly. I've no idea why you had
the problems mentioned, but in general my experiences and those of other
folks I've talked to (at OLS or supercomputing) are not consistent with yours.
In general the requirements for how well a cluster performs is typically with
the vendor, not the parts supplier. If they don't deliver don't pay, exercise
the penalty clause, or protect your investment in some other way.
Last I checked Myrinet was the most popular interconnect after GigE on the
top500 list, I can't imagine it not working well on a large number of
I wouldn't hesitate with recommend myrinet for future clusters, although I
have no personal experience with Myrinet 10G.
More information about the Beowulf