Intel Xeon PHI?

[kriesel]

Originally Posted by chalsall

Have you tried installing Ubuntu instead of CentOS?

Ubuntu won't do, because I'm trying to accomplish the same or very close environment as Ernst settled on for Mlucas development and test after first trying Ubuntu on a 68-core Xeon Phi 7250 also. Ernst's experience with adding numerous packages after using the minimal ISO is also the reason I went big-fat-ISO not minimal (well, attempted to). I have Ubuntu scattered about, as WSL or dual-boot, for occasional test purposes, but Ernst's experience on the Xeon Phi 7250 led him to ditch Ubuntu on that hardware in favor of CentOS, so I'm trying to mimic what he ended up with on the same hardware model.

Also, experimenting in Windows with a single Hynix 2133 Mhz 32GB DIMM (recommended part number) in either 7210 or 7250 gives quite unsatisfactory results. I'm curious what results Ernst may get in CentOS with 6x32GB, especially if he experiments with 1x and 2x before filling all the slots. One DIMM seems to really confuse Windows 10 20H2 (19042.1052) if not the BIOS, about core count and cache, and torpedos prime95 performance and interactive responsiveness. Giving Windows Update another chance now showed it's already up to date. Would like to compare Win & Lin on dual boot on same box for that DIMM effect.

Guess I could give Centos 8.2-2004 a try next for a match to Ernst's version number of 6 months ago. The NICs on both my practice system and my 7250 are known to be cabled and work in Windows.

[chalsall]

Originally Posted by kriesel

Ubuntu won't do, because I'm trying to accomplish the same or very close...

Ummm... ewmayer does some *very* serious work. I would be very surprised if his source code relied on the operating system.

Underlying Compute, on the other hand...

But that's what cross-compilers are for.

[chalsall]

Originally Posted by kriesel

Guess I could give Centos 8.2-2004 a try next for a match to Ernst's version number of 6 months ago. The NICs on both my practice system and my 7250 are known to be cabled and work in Windows.

Could you please expand on what you mean?

I'm confused about the NICs, for example.

[kriesel]

Chalsall, I left you links. Ernst was rewarded for Ubuntu install on 7250 Xeon Phi with what he called "purple screen of death". The cpu and motherboard is unusual hardware which not everything supports well. And attempting wireless right away and CentOS minimal iso created more hassles for him which Paul Underwood et al coached him through, and I think it prudent to avoid those now-known pitfalls. Downloading fat CentOS8.2 now.

Ubuntu onto the practice machine (conventional few-core cpu, motherboard) was not an issue. But Ubuntu onto the eventual Xeon Phi model target was enough of a problem for Ernst to have abandoned it there.

edit: some moderator may wish to move posts 11-14 to the Xeon Phi thread. Or not.

[chalsall]

Originally Posted by kriesel

Downloading fat CentOS8.2 now.

@kriesel. I see you. Please see me.

I always walk with a CentOS 7.9 "Live" USB in the hidden front pocket of my jeans.

There's a reason for this.

[chalsall]

Originally Posted by chalsall

There's a reason for this.

Don't you just all love the "pregnant pause"?

The test came back negative, so we can all then breathe again (or, manage the situation if the test was positive)...

[paulunderwood]

I run Debian Buster on my Phi. Forget multithreading on these beasts. 64 instances is king for mine, although I am running 4 Primo instances with its 256 threads. I don't know how Ernst's code copes.

[ewmayer]

Originally Posted by paulunderwood

I run Debian Buster on my Phi. Forget multithreading on these beasts. 64 instances is king for mine, although I am running 4 Primo instances with its 256 threads. I don't know how Ernst's code copes.

On typical GIMPS wavefront/DC work, I get good Mlucas scaling up to 4 threads (1 per each of 4 physical cores), rapidly diminishing returns beyond that. Large-FFTs get better scaling at high core counts - my current (now ~90%) complete DC run of F30 @64M FFT scales well to 32 threads, maybe ~1.5x added throughput boost from going to 64, which is where I run it. Timing tests of F33 @512M FFT show good scaling to 64, and even a modest added boost from 2-threads on each of 64 physical cores, 128 threads total.

My RAM-install was delayed due to a week-and-half side trip into one added p-1 stage 2 optimization technique for the v20 release (thanks, George). Will do RAM-install this weekend - sorry Ken, will go straight to all 6 DIMM slots filled, no time to mess about with lesser configurations - then if that all goes OK, will suspend my F30 run (and DC on the remaining 4 cores) and embark on v20 p-1 shakedown testing, 17 simultaneous p-1 runs, 4thr/4core each. These will have ~half of jobs find a factor after stage 1, so figuring at most 10 jobs will be doing simultaneous stage 2s, that should allow ~20GB mem-alloc per run. We shall see.

[kriesel]

Originally Posted by ewmayer

DC run of F30 @64M FFT scales well to 32 threads, maybe ~1.5x added throughput boost from going to 64, which is where I run it. Timing tests of F33 @256M FFT

F30 @64M FFT; F33 is ~8x larger; did you perhaps mean F32 @256M there or F33 @512M?

My RAM-install was delayed due to a week-and-half side trip into one added p-1 stage 2 optimization technique for the v20 release (thanks, George). Will do RAM-install this weekend - sorry Ken, will go straight to all 6 DIMM slots filled, no time to mess about with lesser configurations

Any news of DIMMS-added 7250 Phi performance (and any BIOS tweaks necessary to get it) would be welcome.

[ewmayer]

192GB RAM successfully installed and running @2400MHz per the BIOS. But -

1. Runtime for my ongoing LL-DC on cpu 64-67 jumped from 22 ms/iter to 37 ms/iter. The per-iter times for the big F30 DC @64M FFT run on cpu 0-63 are even worse, they went from 66 to 167 ms/iter. I suspect the system may be defaulting to run out of RAM rather than the onboard MCDRAM.

2. Weirdly, both /proc/meminfo and /proc/cpuinfo give "command not found" when run via sudo. When I directly login as root, they instead give "Permission denied".

[paulunderwood]

Originally Posted by ewmayer

192GB RAM successfully installed and running @2400MHz per the BIOS. But -

1. Runtime for my ongoing LL-DC on cpu 64-67 jumped from 22 ms/iter to 37 ms/iter. The per-iter times for the big F30 DC @64M FFT run on cpu 0-63 are even worse, they went from 66 to 167 ms/iter. I suspect the system may be defaulting to run out of RAM rather than the onboard MCDRAM.

2. Weirdly, both /proc/meminfo and /proc/cpuinfo give "command not found" when run via sudo. When I directly login as root, they instead give "Permission denied".

1, is bad news. The access to the DIMMs must be much slower than the MCDRAM -- an expensive lesson. With them, do you have a great 68-instances P-1 machine? See the table on this page. There must be a way to program using mostly MCDRAM much like using caches. More programming info on this page.

2. You have to run cat /proc/cpuinfo. cpuinfo is not an executable.