On 5/28/07, Bill Hart <[EMAIL PROTECTED]> wrote: > These days sage.math is often operating with all 16 cores going flat > out.
This will stop. I'm going to systematically ask all users of sage.math to stop pushing the hardware like this. The *primary* purpose of sage.math is for SAGE software development and collaboration, not individual research (except for the research of William Stein). I will apply for an NSF SCREMS grant, see http://www.nsf.gov/publications/pub_summ.jsp?ods_key=nsf04513 and if the NSF grants it then we can get some serious resources for research computing. > When it gets like this, FLINT slows down dramatically, but MAGMA does > not. This makes timing comparisons meaningless (and slows down > development). My apologies. > I'm wondering if anyone has any suggestions as to why this might be. > How can their code take the same user time whether the machine is > under load or not, and ours takes up to 50% longer. > One problem is that I recently accidentally reniced sshd to -5, which caused all normal user processes to default to -5. I've fixed this. > I've checked the following things: > > 1) Cache hints: We are already using these as these make a noticeable > difference. My theory about this is that when the kernel switches > between different processes, any computation that was underway gets > moved out of cache and the cache gets overwritten by the other > process. Cache hints help by making sure cache coherency is optimised > for the currently active process. > > 2) Nice number: MAGMA runs at a default nice level of -5 as does > FLINT. My fault -- this will switch to 0. > 3) Memory management: a memory manager saves kernel time, but not user > time as far as I can tell. However, the problem I've noted affects > user time. I've experimented with replacing various parts of the > memory allocation and deallocation code with the FLINT memory manager, > but it doesn't seem to affect user time (though naturally it cuts > kernel time). > > 4) gcc compiler optimizations: I use -O3 -fexpensive-optimizations - > funroll-loops > > Does anyone know of anything else? Are there other compiler > optimizations I should use? Is there something else about memory > management that I don't know? Is it to do with the way our library is > linked? Does it have to do with MAGMA having root privileges because > it was installed by root or something? > > There has to be a simple explanation for this. I definitely don't have an explanation, but you're right that there is probably a simple one. That said, it makes no sense at all to do benchmarking on a system that is very heavily loaded. -- William --~--~---------~--~----~------------~-------~--~----~ To post to this group, send email to sage-devel@googlegroups.com To unsubscribe from this group, send email to [EMAIL PROTECTED] For more options, visit this group at http://groups.google.com/group/sage-devel URLs: http://sage.scipy.org/sage/ and http://modular.math.washington.edu/sage/ -~----------~----~----~----~------~----~------~--~---