Reply to post: Re: what about EPYC?

Re: what about EPYC?

But if you look at the work done by the last 4 threads (depending on workload), they very rarely do anything even on a busy system, because the instruction re-ordering is good enough such that four threads will normally consume all the execution units (this is dependent on the mix of instructions in the processes). The way that the execution unit scheduling is done by the hardware thread scheduler means that all hardware threads are not equal, with a decreasing priority, so that for each core, there is one preferred thread, a less preferred secondary thread and so on down to hardware thread 8 , which is the least preferred, and gets almost no time because all the execution units are in use by the other more preferred threads. If you look an the NMON graphs from any of the analyzers, this is very clear.

IIRC, Oracle will accept Power systems running in SMT4, 2 or even 1 as a reduction in number of threads per core for the purposes of their license calculation. We run our Oracle systems at SMT 4, and even at that, the last hardware thread is normally showing little work actually scheduled.