On Tue, Sep 20, 2005 at 08:28:11PM +0200, umaxx wrote: ... > 2. Is it better to test performance with disabled Hyperthreading?
In general, yes. > perftest/rdma_lat is using get_cycles(). get_cycles() uses internal > the rdtscll() function. Correct me if I am wrong, but this could be a > problem with a hyperthreading enabled cpu? Because it's not clear which > cpu(-core) is used by rdtsll() at which time? The time stamp counters > between processors are not synchronized, and therefore the timing > measured will be incorrect. I don't know if this is a problem in practice. It potentually could be. > Maybe something like cpuid() is missing - to set/get the wanted? I suggest using taskset(1) to bind a test program to a logical CPU if it really is a problem. The tests have fairly short run times and it's not worth making them more complex to deal with this. Someone else might take a shot at answering the other two questions. hth, grant _______________________________________________ openib-general mailing list [email protected] http://openib.org/mailman/listinfo/openib-general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
