Author: Vincent Diepeveen
Date: 17:04:40 07/15/03
Go up one level in this thread
On July 15, 2003 at 10:25:25, Gerd Isenberg wrote: >Hi Vincent, > >Thanks, for the source code. I will try it during the next days... >How long will the test run? > >Another interesting thread about this issue: > >http://www.sandpile.org/post/msgs/20003914.htm >http://www.sandpile.org/post/msgs/20003911.htm > >with some comments from Paul Hsieh: > >http://www.sandpile.org/post/msgs/20003912.htm > >It would be interesting to compare the results from lm-bench with Oliver's and >your's on different hardware. > >Regards, >Gerd Note that the test i shipped you allocates 1 big memory block in OpenMP. For clusters having openMP or cc-NUMA machines having openMP it is better to use my latencyC.c test which allocates locally a hashtable at each processor and with openMP connects it to each other. Then they read randomly from each processor. I already have enough results from different supercomputers and PC's at both testsets to know what is good in at SGI supercomputers and supercomputers in general and what is not so good there. However i would love to get from both testsets some results back from opterons and dual opterons. that would be real cool. don't have any yet. we know the sequential latency is 40-50 ns claimed. so let's see what the time is to get 8 bytes at a random adress out of memory. I'm hoping for < 150 ns or something of course. I fear it's only slightly better than the 280ns that P4/K7 delivers single cpu. Like 20% at most. So like 200 ns.
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.