SG BOINC Pentathlon 2021

[dannyridel]

Well, it looks like city run is PrimeGrid!!!
PPS Sv will get a huge boost this time certainly.

[SuicideCabbage]

Are we participating in this first throw? I'll start moving cores over, but want to make sure I'm not throwing alone.

[Dirk Broer]

I'm running MIP as well -I am always running it. I suggest everybody joins in the first throw, as there's only the Marathon (SiDock) that might be interfering, and that lasts 14 days.
Set all your other WCG projects temporary on hold, and run MIP for 24 hours to see what we can get. If we do well, let's do the 2nd one too.
06 May 2021 00:00 - 07 May 2021 00:00 UTC - 1st throw
08 May 2021 00:00 - 09 May 2021 00:00 UTC - 2nd throw

I also suggest that we run PrimeGrid (mainly) as GPU project, the PPS Sieve (PPS SV) pays very good and runs light. You might consider Wieferich & Wall-Sun-Sun (WW) too.

[SuicideCabbage]

All of my big guns are pointing at MIP as we speak.

I know nothing of PrimeGrid so the sub-project recommendations are very helpful. Does it fully utilize a GPU or should they be ran multiple units per card for maximum throughput?

[Dirk Broer]

PPS Sieve and Wieferich & Wall-Sun-Sun are both GPU sub-projects that bring lots of credits when crunched with a recent GPU.
A GTX 1650 scores on the same level as an older Tesla, like the ones you get when you use Google colabs. A GTX 1660 Ti is even better: it flies through them.
To say nothing about the RTX 2000 and 3000 series. The same should hold true for the latest two series of Radeon cards.
You can possibly run more WUs on a recent card with loads of RAM, but that's not within my budget.

[dannyridel]

Also, on CPUs with AVX running LLR is highly recommended. Run LLR with hyperthreading closed. It decreases throughput. Or just set BOINC to use 50% of CPUs in computing prefs.
Wiefierich and Wall-Sun-Sun favors Turing GPUs, while PPS SV is the overall best choice. TRP is a sweet spot too for CPU projects.
I'm going to dump my bunker when I get home, my main computer is crunching with network activity suspended. PPS SV can be bunkered too!

[vaughan]

@SuicideCabbage you can run multiple GPU tasks using an app_config.xml
On my faster cards I run 3 tasks simultaneously and make sure you leave 1 CPU core free to feed the GPU else it gets throttled.

<app_config>
<app>
<name>pps_sr2sieve</name>
<max_concurrent>3</max_concurrent>
</app>
<app_version>
<app_name>pps_sr2sieve</app_name>
<plan_class>cudaPPSsieve</plan_class>
<avg_ncpus>0.01</avg_ncpus>
<ngpus>.333</ngpus>
</app_version>
</app_config>

[vaughan]

The SiDock tasks really take too long and as the application doesn't have check-pointing it is quite a worry that hours of CPU time and electricity could be wasted.

[Dirk Broer]

Also, on CPUs with AVX running LLR is highly recommended. Run LLR with hyperthreading closed. It decreases throughput. Or just set BOINC to use 50% of CPUs in computing prefs.
Wiefierich and Wall-Sun-Sun favors Turing GPUs, while PPS SV is the overall best choice. TRP is a sweet spot too for CPU projects.
I'm going to dump my bunker when I get home, my main computer is crunching with network activity suspended. PPS SV can be bunkered too!

That's why I advise to run PrimeGrid as GPU project, so you can use your hyperthreading for the other projects in the Pentathlon. And while Wieferich & Wall-Sun-Sun may favor Turing GPUs it does quite well on the 1600 series t00.

[Dirk Broer]

Cross-country will be Einstein@Home