Need mprime configuration advice for P-1

[Mark Rose]

Originally Posted by kriesel

Welcome to the party. There's plenty to do. It's now at ~85.5M and there are a lot of them.
Try one worker and two workers, and use whatever gives better aggregate throughput. Four workers would either leave 3 stage 1 oversaturating one stage 2 at a time, or divide the memory into 2 stage two 15 GiB workers which is below George's threshold of 24 GiB.

I took a few hundred assignments for the first four Skylake systems. I've got some DC work to finish up with one worker on each.

I configured them with 2 workers, 2 cores, with a single high memory worker. If stage 2 is balanced correctly by mprime, each stage 2 run should have 30 GB to use while the other work runs a stage 1.

[Mark Rose]

Tangentially related: When mprime prints out Time numbers doing P-1, are lower or higher numbers better?

I'm running it on an 8-core laptop and see lower numbers when using 6 cores versus 8, which is counter intuitive. Or is the system actually slower running more cores?

[paulunderwood]

Originally Posted by Mark Rose

Tangentially related: When mprime prints out Time numbers doing P-1, are lower or higher numbers better?

I'm running it on an 8-core laptop and see lower numbers when using 6 cores versus 8, which is counter intuitive. Or is the system actually slower running more cores?

Could this laptop of yours be TDP limited? That might be what's happening.

[Mark Rose]

Originally Posted by paulunderwood

Could this laptop of yours be TDP limited? That might be what's happening.

Yes. The six active cores are throttled to 3.8 GHz (base 2.4 GHz, max turbo 5.3 GHz)

But I'm a bit surprised that 8 cores at even lower clocks wouldn't deliver more throughput, if lower "Time" numbers are better. Are lower "Time" numbers better?

[paulunderwood]

Originally Posted by Mark Rose

Yes. The six active cores are throttled to 3.8 GHz (base 2.4 GHz, max turbo 5.3 GHz)

But I'm a bit surprised that 8 cores at even lower clocks wouldn't deliver more throughput, if lower "Time" numbers are better.

Do the math: core count times GHz, and compare.
"
I think lower "time" is better.

[Mark Rose]

After playing around a bit, at least on this i9-10885H system with dual channel memory, a single worker doing P-1 using 4 cores gives the lowest time numbers. Configuring two workers to maximize the RAM usage is slower. Interestingly, in this configuration, stage 1 does not thermal throttle. Stage 2 does.

[Prime95]

Be careful. Measure the change-in-B2-completed-values divided by time. Using fewer cores may result in a lower target B2 value.

[chalsall]

Originally Posted by Prime95

Be careful. Measure the change-in-B2-completed-values divided by time. Using fewer cores may result in a lower target B2 value.

Agreed.

I have found that when using most COTS kit it is relatively trivial to bring the thermals up high. Without much effort. Burn baby, burn.

Measuring stuff is just what we were trained to do.

Correct?

[Mark Rose]

So all my Stage 2 testing was done on the same exponent in progress.

But I could definitely see running a second worker impacting the timing.

I'll let it run single worker, 4 cores, on a new exponent, to see what kind of timing I get.

[chalsall]

Originally Posted by Mark Rose

So all my Stage 2 testing was done on the same exponent in progress.

Thank you for this work.

Is it safe to assume you are currently working within an Intel compute space?

[Mark Rose]

Originally Posted by chalsall

Thank you for this work.

Is it safe to assume you are currently working within an Intel compute space?

Yes. I don't have any AMD systems at the moment. Haven't bought a new system personally in many years, since the five i5-6600 systems.