Sgt. Joe,

As per threat/promise, back again :-)

I had been reading your messages as if you were saying there was no change of workload pattern, but as soon as I thought about how I set my systems up I realized that might not be the case here! So I have a question -- before your MCM1 tasks went on a go-slow, were you running a mix of MCM1 and SCC1? If so, I think we have an answer to your slow-down :-)

I've just run a little test on a 6 core/12 thread Ryzen which normally runs 3 MCM1 and 2 SCC1, along with a 4-thread MilkyWay N-body task. (The average CPU time for Ovarian MCM1 tasks was about the same with or without the SCC1 tasks.) Here are some MIPS calculations based on observations made with the Linux perf stat utility to examine aspects of performance for a single MCM1 task whilst running various numbers of MCM1 tasks and no other BOINC work...

        One MCM1 task      12,880 MIPS
        Three MCM1 tasks   12,113 MIPS
        Five MCM1 tasks    10,887 MIPS
        Eight MCM1 tasks    8,377 MIPS

Note the significant decline as the number of threads used approaches and passes the number of actual cores :-)

I suspect the effects might be even more marked on older hardware with less FP capability and L3 cache...

As an aside, I also noted that MCM1 tasks seem to produce a lot of page faults! I suspect that's a side-effect of bouncing backwards and forwards across a 40+MB dataset, and I don't think there's much that can be done to alleviate the stresses that'll put on page tables and such like... (How much of a performance hit might result is unclear.)

Cheers - Al.

P.S. As for the other quest -- parameters affecting run time -- I've not found any evidence of major differences as yet (and somehow I don't think I will find any...)