From: Andrew T. <hab...@us...> - 2003-02-04 14:01:15
|
On Tuesday 04 February 2003 04:23, Ingo Molnar wrote: > On Mon, 3 Feb 2003, Andrew Theurer wrote: > > In case anyone is interested, here is SchedD7, with and without HT bi= ts, > > compared to SchedD7 minus HT bits plus numa-ht topology, compared to = a > > run with E2: > > > > The system is a dual P4 xeon, serverworks chipset, kernel is 2.5.59. = I > > ran "kernbench" with -j2, -j4, and -j8. Kernbench result is an avera= ge > > of 10 runs. The kernels are: > > thanks for the analysis - looks like the -E2 scheduler got the best > numbers in every benchmark, the difference is especially visible in the > -j2 test, where the difference between schedD7-noht and schedE2 is more > than 30%. But -E2 even beats the best NUMA-HT scheduler > (2.5.59-D7-numaht2) by 8%. I think it's going to be hard, if not impossible to beat E2 with a numa-s= ched=20 for HT. The approaches are a little different, and E2 just may be the mo= re=20 efficient approach. The numa scheduler tries to find the best cpu right = when=20 the new process is exec'd, but does not do an "active" load balance that = the=20 D7&E2 patches do. The sched_best_cpu() from numa_sched can take some tim= e to=20 complete, while active load balance can help when some tasks exit and a H= T=20 type imbalance exists. Although the HT-numa way is a good improvement ov= er=20 stock 2559, it just doesn't approach the performance of E2. Also, FYI, last time Michael tested with D7, he had some weirdness with a= =20 "real" NUMA system. Hopefully we can reproduce that soon on E2 and track= =20 down any issues. =20 -Andrew Theurer |