BOINC Pentathlon 2024

[leadeater]

5 minutes ago, Blasty Blosty said:

Ah I see, so similar to Folding @ Home?

Pretty much yea. BOINC Client App is a GUI front end to run one or many projects like F@H so there's a little more to it but pretty much the same thing with more steps.

[Blasty Blosty]

Just now, leadeater said:

Pretty much yea, BOINC Client App is a GUI front end to run one or many projects like F@H so there's a little more to it but pretty much the same thing with more steps.

Thanks for the info, I might take part

[rkv_2401]

7 hours ago, Blasty Blosty said:

Ah I see, so similar to Folding @ Home?

Yup

[WhitetailAni]

Ooh nice! Maybe I'll bring back the Pentium 3 BOINC box - it's got an 800 MHz Coppermine P3 now instead of a 500 MHz Katmai.

 

I think it took three days for one WU last time?

[Ghett Klapson]

cli users beware, don't accidentally run `boinc` in a directory that has a dir named `projects`. IT WILL DELETE ALL FILES IN THAT DIRECTORY WITHOUT WARNING!

[dogwitch]

8 hours ago, Ghett Klapson said:

cli users beware, don't accidentally run `boinc` in a directory that has a dir named `projects`. IT WILL DELETE ALL FILES IN THAT DIRECTORY WITHOUT WARNING!

old known bug

[Neptune_cj]

Will there be any prizes offered like in the month of folding at home?

[Lightwreather]

22 minutes ago, Neptune_cj said:

Will there be any prizes offered like in the month of folding at home?

I don't we have had prizes for previous BOINC Pentathlons at all

[leadeater]

3 hours ago, Neptune_cj said:

Will there be any prizes offered like in the month of folding at home?

 

3 hours ago, Lightwreather said:

I don't we have had prizes for previous BOINC Pentathlons at all

 

Prize are a little complicated since the Pentathlon rules don't really allow it, not something we noticed much before but it's probably better to not offer them.

 

Quote

The organizers do not approve of systems that are based on monetization instead of enthusiasm

 

[dogwitch]

1 hour ago, leadeater said:

 

 

Prize are a little complicated since the Pentathlon rules don't really allow it, not something we noticed much before but it's probably better to not offer them.

 

 

i be happy with a prize of a real  pic of the Almighty lead eater.

[leadeater]

1 hour ago, dogwitch said:

i be happy with a prize of a real  pic of the Almighty lead eater.

[dogwitch]

5 minutes ago, leadeater said:

this is what i think lead really looks like!

[Blasty Blosty]

For my contribution prize I would like a 4090

[Dutch_Master]

4 hours ago, Blasty Blosty said:

I would like a 4090

Can't do that, settle for a 4070 (<-link!) instead?

[Arca Ege]

4 hours ago, Dutch_Master said:

Can't do that, settle for a 4070 (<-link!) instead?

No there's a 4090 too

[Windows7ge]

On 4/28/2024 at 11:56 PM, leadeater said:

@Windows7ge @Gorgon

 

I don't know about this year due to circumstances but maybe next year. o7

[AamirA]

My hardware isn't the best, but something about these boinc pentathlons really makes me exited. I like the strategy of it compared to just raw computing power (even though it's mainly just that). I had so much fun last year. But why do they have to do these competitions right when it gets to 80 degrees!? I guess it's part of the fun

[SpaceRazor]

Is this something I need to sign up before it starts? I already use Boinc on team LTT, but I want to make sure I am contributing correctly for the event. Thanks.

[leadeater]

1 hour ago, SpaceRazor said:

Is this something I need to sign up before it starts? I already use Boinc on team LTT, but I want to make sure I am contributing correctly for the event. Thanks.

You only need to check this topic first post or the BOINC event website and make sure you are doing work for the correct active project(s). Sign up is for the team itself and that is done.

[Blasty Blosty]

12 hours ago, Dutch_Master said:

4070 (<-link!)

Good enough, I'll take it

[Blasty Blosty]

8 hours ago, Arca Ege said:

No there's a 4090 too

omg even better thanks guys

[TheLANguy]

It's been quite a while since I used BIOINC, and I never knew about the BOINC Pentathlon!

Just got it setup running PrimeGrid. 

Can you use the same account on multiple PCs to run more WU? Or is there a different/better way?

[leadeater]

6 hours ago, TheLANguy said:

Can you use the same account on multiple PCs to run more WU? Or is there a different/better way?

Yes use same account

[porina]

Marathon project has been announced. It is Primegrid's Cullen subproject. For those not familiar, it is similar to running Prime95 large FFT so prepare yourself!

 

Rough optimisation guide:

The tasks currently use about 21MB of data (see edit at bottom for updates) and generally run fastest if you run multiple tasks in total not exceeding the L3 cache size of your CPU. If one task exceeds the L3 cache of your CPU, run 1 task. The number of threads per task can be controlled in Primegrid settings. If you have multiple different systems, you can use more than one venue. Generally speaking run it only on real cores so don't fill up all threads.

 

e.g. on my 7800X3D with 96MB of L3 cache, in theory I should run 4 units at the same time for maximum performance.

 

This can be checked by running a benchmark in Prime95

 

Try using the settings above as a starting point. 2688 is the current largest FFT size used by the project. This may go up during the challenge! Note the software can use different optimal FFT sizes depending on CPU architecture. If you get no result, try lowering the minimum size a bit to find the nearest lower FFT size.

Tick the 2nd check box, as the project uses LLR. It doesn't seem to make much difference so probably doesn't matter if you forget.

Uncheck "Benchmark Hyperthreading" as generally it doesn't help but still increases power consumption on Intel CPUs. You can try it anyway if you want to see what happens.

"Number of workers" is the number of tasks to try at the same time. On my system it defaulted to 1, 2, 8. You can edit it for other divisions of your core count. I added 4 here, as that is interesting.

You can reduce the time to run at the bottom to the minimum 5 seconds to make it a bit faster.

 

Quote

Timings for 2688K FFT length (8 cores, 1 worker):  0.79 ms.  Throughput: 1264.35 iter/sec.
Timings for 2688K FFT length (8 cores, 2 workers):  1.55,  1.54 ms.  Throughput: 1297.16 iter/sec.
Timings for 2688K FFT length (8 cores, 4 workers):  3.05,  3.02,  3.03,  3.04 ms.  Throughput: 1318.40 iter/sec.
Timings for 2688K FFT length (8 cores, 8 workers):  7.61,  7.59,  7.52,  7.37,  7.60,  7.40,  7.44,  7.57 ms.  Throughput: 1065.11 iter/sec.

You should see output something like the above. Basically you want the biggest number after throughput. As theory predicted, 4 tasks (4 workers) at the same time gave the highest throughput, but note it isn't that much faster than one task using all the cores, about 4%. You may chose slightly lower throughput if you want shorter individual task runtimes, which could be handy at the end if you want to squeeze out a unit before the deadline.

 

Note benchmark numbers can't be directly compared at different FFT sizes! They get lower as FFT size gets bigger.

 

Special considerations:

• For CPUs with split L3 caches (Zen 2 and earlier with 6 or more cores, Zen 3 and newer with 12 or more cores) you want to consider each CCX as its own CPU initially. Generally you'll probably run 1 task per CCX if the cache is big enough. If not, one task per whole CPU. 3rd party software like Process Lasso can help prevent Windows messing up scheduling but I can't advise the exact settings required.
• Similar to above applies if you have NUMA or multi-socket systems.
• Because the software doesn't generally benefit from HT/SMT, if you have it enabled on your system you should set the BOINC client to use the % of cores vs threads. e.g. if you have a 8 core 16 thread system, set BOINC client to use 50% of CPU cores. Alternatively you can turn off HT/SMT and run 100%.
• I don't have experience with hybrid Intel CPUs. You'll have to test if using E cores helps or hinders. I'd guess only using P cores is probably optimal.
• If you have Intel mesh cache CPUs (e.g. Skylake-X, Cascade Lake-X and similar Xeons) count L2 + L3 instead of L3 alone. These are heavily weighted to L2 and it does help. On most other CPUs L2 is insignificant.
• If you don't have enough cache for 1 task, it will use ram and therefore ram performance will impact overall performance. Bandwidth generally helps more than timings.

 

Edit: as expected the FFT size is increasing over time.

Max FFT size and associated data size per work unit:

2 May AM 2688k - 21MB

4 May PM 2880k - 22.5MB

[leadeater]

@porina Any suggestions for running the above benchmark on Linux server cli only with multiple sockets, AMD and Intel. Hopefully Primegrid Linux will nicely respect NUMA nodes and L3 boundaries but maybe it won't.