Heath Volmer wrote: > Now running linux on the same box; When I run two instances of > mprime, both of them run at 100%. Is it fair to assume that I'm > getting full output out of both of them? How do I tell? >
a HT P4 is best off running ONE thread of mprime thats affinity locked to one of the virtual CPUs. the HT threads are OK for integer processing, but two pipeline optimized FPU bound threads will just trip up each other and run slower than one at a time. the hyperthreads on those cpus have only one set of execution units (address cruncher, integer adder, integer mul/div unit, fpu) but two sets of registers, and two execution pipelines and instruction decoders. the P4's pipeline is long and narrow, and execution units are often idle during any given clock... by running two threads at once, the cpu can achieve better overall utilization of its hardware. While the usual gain of a mixed integer workload is around 1.3X, we have some java network middleware workloads that achieve 1.7X better throughput on a dual xeon HT system with HT on and an optimized worker thread count. _______________________________________________ Prime mailing list [email protected] http://hogranch.com/mailman/listinfo/prime
