Competition side effects?
9 May 2019 15:38:15 UTC
Topic 218828
(moderation:
I've noticed for some days now that my pending GPU tasks count is climbing steadily, well out of the normal range of variation (I log it daily as I add in an estimated credit for pending work to smooth my monitoring). To be specific my pending GRP task count has climbed from 653 on April 30 to 1571 just now. I've also noticed RAC dropping--as the rising pending has subtracted enough credit to notice a little.
Clicking about in the middle of my pending heap of tasks, I detected a number of machines which acquired a large number of Einstein GPU tasks starting on or about May 2, and in some cases continuing to the present, without returning any work for some days.
As someone here had https://einsteinathome.org/content/radeon-vega#comment-171200 mentioned a Pentathlon in a Radeon Vega thread, I looked around and found:
https://www.seti-germany.de/boinc_pentathlon/start.php
The specs on that page state that the date range for the Einstein GPU "event" is from midnight UTC May 11 through May 16. So if competitors in this event are responsible for the effects I have seen, I supposed the pending elevation will largely reverse within the week (and actually that my RAC will balloon above the legitimate level before settling).
Any comments from pentathlon participants or those in the know?
Many who compete in the
)
Many who compete in the Pentathlon "bunker" tasks. (Download and crunch them early, then report them after the start of the event) Expect to see a flood of these then.
You are correct in your
)
You are correct in your assumption as to why your pendings are climbing. All should be well within 30 hours or so from now.
Yep, thats most likely the
)
Yep, thats most likely the reason. Even several days prior to the official announcement of E@H as the Cross Country event (there were several projects as options) I was downloading tasks and turning off networking so they could all be 'completed' during the event time frame.
The Pent will bring many new users to the project and BOINC in general. It's how I started from only running FAH. Even mega projects like WCG had a 33% boost in production on the 1st day. YoYo has about tripple its normal production and NFS has 3-4x on it's single event days. Most chosen projects like the increased production from the event and I believe they are all asked ahead of time to opt in/out as a possible project.
Beginning in 11 hours people will start to release their tasks and the validators will be put to the test. :) E@H has been a project in the Pent and FB Sprint in years past and have handled it pretty well.
The competition will be
)
The competition will be updating a page giving team credits at Einstein at :
https://www.seti-germany.de/boinc_pentathlon/statistiken/challenge.php?challengeid=4&lang=e
I sure hope the E@H folks
)
I sure hope the E@H folks that keep the queue filled don't just go home for the weekend and miss all the fun. Already looks like they're having trouble supplying enough WUs. I know I could use a lot more than I have.
In 12 hours the dam will break.
archae86 wrote:The
)
If you're curious about the ridiculous amount of damage we're about to inflict, follow that link starting in ~91 minutes.
1st update today was close to
)
1st update today was close to double any other day with another update to go. But very little work to go around.
archae86 wrote:I supposed the
)
With the Einstein-related component of the Pentathlon closed a few minutes ago, I can say that my pending dropped from 1,914 the minute before the competition began to 846 a couple of minutes after it ended. My RAC recovered but did not "balloon" very much above the real level.
The competitors who were my quorum partners clearly did some gamesmanship in timing their reporting of units. The pending units vs. time graph has some distinct non-random features.
My end-state pending is enough higher than my recent average of about 650 to make me suspect that not all of the tasks taken up by competitors have been returned. I wonder whether the participants will eventually process and report them, abort them, or just abandon them to time out.
I can't speak for everyone,
)
I can't speak for everyone, but i'm crunching all of mine. 9 left to go.
archae86 wrote:I wonder
)
Since quite a few of the 'contestants' seem to be rather fanatical about the 'competition', it would make no sense for them to keep any excess around where it could interfere with the ability to download the maximum possible for the next 'event' to follow this one. I would expect all excess to be immediately aborted.
You could hope that they would have been courteous enough to plan for the end of the current phase by calculating what was needed with a small safety margin, thereby limiting the unnecessary waste. After all these guys seem to have no trouble calculating how to squeeze the max credit possible by all sorts of 'dubious' methods like bunkering :-). Human nature being what it is, I would predict that many would not have considered what might be courteous but would just have downloaded the max with the intention of immediately aborting the high numbers of 'leftovers' close to the end.
I replenish my caches under script control approximately 4 times per day. Prior to requesting work, every host is synced up with a full set of 'LATeahnnnn' data files from a server machine on the LAN so that only 1 copy of each brand new file is ever downloaded. The server gets the new file immediately and every other host gets their own copy from that. Because downloading of tasks is strictly controlled, I can see exactly what happens for any host I choose whenever it 'takes a drink' :-).
I just picked one host at random and looked at what it received around 6.00am this morning and again around noon (today, Thursday, where local time is UTC+10). Those two times were about 4hours before and 2 hours after the end of the event for E@H. For those two 'top-ups', the host received 33 resends out of a total of 33 tasks requested. I think that's pretty conclusive evidence of a 'big dump' of excess stock :-). The host I looked at has a HD7850 GPU and completes about 70 tasks per day. For the whole of yesterday, (Wednesday, local time) it requested 69 tasks of which precisely nil were resends :-).
Those 33 resends covered all previous data files from LATeah1049V to LATeah1049ZA - that's 6 data files. I'm running 107 GPU hosts so (assuming similar behaviour on all machines) that indicates something like around 600 data file downloads that would have been forced on me if I hadn't set up my data file caching system. It annoys me when (seemingly) little consideration is given to the 'collateral damage' that such actions can have on the many ordinary volunteers just trying to support the project for the longer term. In fact you sometimes see an expression of satisfaction at a 'job well done' if the project happens to stumble under the onslaught :-).
Cheers,
Gary.