SG BOINC Pentathlon 2022

[dannyridel]

But what about Einstein? I use my GPUs for that, as (most of my) GPUs offer no real advantage over CPUs with NF.
But when we forget about Einstein altogether (least points for us so far) I could live with that.

For NF, we just have to beat the 610k day.

[Dirk Broer]

We did beat the 610k day, set NF to 'suspend' till the last throw. We need to beat the 610k day yet again, as the 3rd best score counts.

Edit: I take it that you guys/gals had the same experience with Universe and Einstein that I had (not being able to load up), and went on crunching Numberfields for the heck of it....

[dannyridel]

Lol that's me indeed. At this rate I might start crunching Einstein with my CPUs as well (the server seems to be better than universe). We could maybe try to get Italy back with a stronger last day?

And we need to defend for P17 from LITOMYSL.

[vaughan]

So what is actually the problem with the Universe@Home server? I realize the Pentathlon has more users / computers running but how much more than usual? The tasks appear small in size so its not a bandwidth problem, so what is it?

Similarly, Einstein project seems overloaded but their tasks are huge so lack of bandwidth explains its problems.

PrimeGrid's server(s) coped with the influx of users the other day. What do they do differently?

Also, Numberfields appears to be handling the load.

[SuicideCabbage]

I had over 4000 complete Universe tasks at the start of the day (UTC), I managed to get ~250 or so up, ended the day with about 4200 complete. The servers are an absolute joke, it was cute at the start, a power overwhelming look what we did kinda thing, but it's come to the point it's just infuriating. NumberFields fell on their face the first two throws, first one being down for at least 6.5h before I gave up and went to bed, the 2nd day down for about an hour and a half, but they managed to get their crap together. Even during non-bonus days Universe has been a disgrace, my 64-core machine ate through it's 2,304 allotted units that took me two days to get after the first bonus day purge thanks to timeouts and no work messages. BOINC refuses to ask for work with X amount of uploads in progress and I have been unable to get work on that machine for five days now, it's been running Einstein or NumberFields just to keep it warmed up. The five days without work is ~1.2mill in credits lost on that machine alone, and it's not much better across the entire grid. If I had to guess, across all the cores I have set to Universe as a primary project, I've had 14 CPU MONTHS of idle time just since the start of the Pent. (Universe idle time, don't worry, they are doing something) I've been trying to purge the completed work on the 64-core machine since the SECOND bonus day! The issue is they need 6 or 8 files uploaded per workunit, and when the TCP handshake is larger than the data payload itself, their 56K modem is just hammered to death by requests. I entered the day needing to make 24,000-32,000 uploads to get rid of the work I had completed on hand, their network is trash for sure, but zipping the output files into a single file for upload may bring the load down to manageable levels. If Universe is ever selected for the Pent again, which I wouldn't be upset if they were not, I'm going to run anything larger than a 6 core machine as VMs, limited to four cores. At least then the timeout lottery will shine on a few of them, and part of the processor will remain busy. If not by next year, then by the year after I will have a 256 thread EPYC server, if I had that today, it would be next years Pent before it even managed to report the work from this one.

[Dirk Broer]

I think we should go full out on Numberfields till we pass the 618k of our 1st throw, so it becomes our 3rd best. After that point is reached we can go full out on Einstein, as our upload queues (at least those of SuicideCabbage and me) for Universe are already filled to the brink.

[vaughan]

Summary from Seti-Germany:

Javelin Throw (NumberFields@home) AMD Users secured 13th position.

City Run (Einstein@Home) Chinese Dream could only just keep #17. AMD Users came dangerously close in the end, but it still was not enough, and so the hardware faction has to settle for #18.

Obstacle Run (Universe@Home) With #11, Meisterkuehler took their best rank in a discipline this year and thus surprised OCN (#12), BOINC.Italy (#13), and AMD Users (#14), costing them a few points in the Overall Standings.

Cross Country (PrimeGrid) 11th only 1.5 million points behind Ukraine.

[dannyridel]

Still-awesome work guys for 12th! Congrats to all on that last Javelin NF push-it really worked out!