LHC@home

I happened to be running all CMS today, due to problems with the other types, and found that only 12 work units would run on my Ryzen 1700, with 15 cores available. Apparently it is not a memory shortage (I have 32 GB), and there was no error message indicating why any tasks would not run.

So I paused 4 of the CMS, and immediately 7 others started (3 Theory, 3 LHCb and 1 CMS), so that all 15 cores were utilized. So it seems that LCH is imposing a 12-task limit on CMS. Note that I have "Max # jobs: No limit" set in my preferences. I don't recall ever hearing about this.

15 CMS-tasks would allocate 34.92 GB of memory and depending on your BOINC memory settings, it probably will not run more then 12 tasks, I suppose.

Yes, that is probably it, even though I did not see any tasks waiting to run, and no memory error messages. Also, when I suspended the four CMS and got the others five to run, they should have used at least as much memory as the suspended ones. But there may be a difference between what each work unit reserves for itself as a requirement, and what it actually uses. I must have bumped up against some limit.

I tried adding more memory, but with four memory modules the speed drops down from 2666 MHz to 2133 MHz. It seems to be a common problem with some of the Ryzen motherboard/memory combinations. So the solution for me is to build a Coffee Lake machine. It will have only twelve cores, but each core will run faster than the Ryzen 1700, so the output should be about the same. Also, I will try to get a little above 32 GB memory if possible, for some extra margin. I don't want to run into the situation where the machine freezes due to lack of memory, as I have seen before.

Thanks for your input.

I tried adding more memory, but with four memory modules the speed drops down from 2666 MHz to 2133 MHz.

If you have different RAM-Modules, you can see this in CPU-Z.
If Yes, the lowest RAM-MHz is used.

If you have different RAM-Modules, you can see this in CPU-Z.
If Yes, the lowest RAM-MHz is used.

They were all identical Patriot 2666 MHz modules (16 GB each). It is a problem with the AM4 motherboards; they tend to be difficult, even with the latest BIOS. The Intel might (possibly) be better.

grrr...
then you can only hope for a BIOS update.
The Ryzen is my favorite for the future!

I tried adding more memory, but with four memory modules the speed drops down from 2666 MHz to 2133 MHz.

Do you think that the memory speed drop has a noticeable impact on the speed/efficiency/etc. of boinc/cern calculations?

grrr...
then you can only hope for a BIOS update.
The Ryzen is my favorite for the future!

It is irrelevant now. I will use only 32 GB at 2666 MHz (two modules) in this machine, and run it on WCG. I will also be building a Ryzen 2700 later, which will have new motherboards and probably have improved memory compatibility.

I tried adding more memory, but with four memory modules the speed drops down from 2666 MHz to 2133 MHz.

Do you think that the memory speed drop has a noticeable impact on the speed/efficiency/etc. of boinc/cern calculations?

I noticed it a bit - maybe 5%, though I did not really try to measure it. It is not that big of a deal, but I will be building new machines anyway. It is more a case of allocating the right machine to the right project, and the amount of memory affects that.

[The Ryzen is my favorite for the future!

I like it a lot. I originally had an i7-4790 on LHC, and the Ryzen 1700 was originally on WCG. But when I tried the Ryzen here, I found that it was better per core (and of course more cores), at least on Theory, and I think the others too. But it does well on a lot of other projects too (except for Rosetta for some strange reason), so it can be used in a lot of places.