On Mar 11, 12:58 am, Alex Ghitza <aghi...@gmail.com> wrote: > On Thu, 11 Mar 2010 00:41:16 +0000, "Dr. David Kirkby" > <david.kir...@onetel.net> wrote: > > > > > > > > NUM_THREADS=10 > > > in makefile. Then I issued "make ptestlong", only to see that cddlib > > > was reinstalled. Doctesting is still running in parallel using 10 > > > threads. > > > I've not tried a parallel ptestlong myself. I'll see what happens on one of > > my > > machines. I can't see why cddlib would be reinstalled. > > > It is probably worth using far more than 10 threads on 't2'. Although it > > has 16 > > core, there are 128 hardware threads. I've convinced myself that one needs > > to > > use at least 256 threads to get optimal performance. In fact, going to 1000 > > threads is very slightly quicker. > > > I think I posted some benchmarks I'd done before on 't2', and quite simply > > setting threads=cores does not give anywhere near the performance possible. > > Of > > course, you don't want to bring 't2' to its knees, but it is not heavily > > used. > > We have to be careful here and experiment some more before > recommending a number. A couple of weeks ago I tried make ptestlong > with 100 threads on t2 and had to eventually kill everything in sight > manually to avoid disaster :). > > Best, > Alex
What made you believed disaster would strike? The machine does not have a lot of RAM (only 32 GB), so I could imagine that memory exhaustion could be an issue. One would have to look at the scan rate with 'vmstat' or similar. Be aware, 'top' is not useful on modern versions of Solaris. The ZFS file systems will grab tons of memory, so care neds to be used in interpreting any tool which aims to measure free RAM. Solaris will basically make use of whatever RAM there is, so the concept of free meory is not very useful. Also, load average is not a particulary useful measure. I must admit, I don't fully understand how to analyse the T5240 is coping with whahtever load is on it. Dave -- To post to this group, send an email to sage-devel@googlegroups.com To unsubscribe from this group, send an email to sage-devel+unsubscr...@googlegroups.com For more options, visit this group at http://groups.google.com/group/sage-devel URL: http://www.sagemath.org
