Message boards :
Number crunching :
Problem with tiny cache in MW
Message board moderation
Previous · 1 . . . 5 · 6 · 7 · 8
Author | Message |
---|---|
Send message Joined: 2 Jan 08 Posts: 79 Credit: 365,471,675 RAC: 0 |
I am happy here if work units going to be 4 times bigger, as told on the front page. Thanks Travis. |
Send message Joined: 26 Jan 09 Posts: 589 Credit: 497,834,261 RAC: 0 |
|
Send message Joined: 24 Dec 07 Posts: 1947 Credit: 240,884,648 RAC: 0 |
There we go. A bigger cache! WOOHOO! Let's see, 24 wu's @ 190sec avg with 2 GPUs = 38 min! LOL, but much better than previously where 24 wu's @ 48 sec avg with 2 GPUs = 9.6 min! A step in the right direction. Mind you I haven't seen any web site slowness or wu unavailability that I've been concerned about. Mind you on my old P4 with the 3850 I now have a 2hr cache! And on my dual core CPU only machines the cache is going to blow out to 32hrs for 12 wu's and that is without taking into account resource share....so much for fast turn around times. Wow! Here's a thought, the project may need to start considering checkpointing! ;) Travis, time to go to a GPU only project? Live long and BOINC. |
Send message Joined: 18 Sep 09 Posts: 4 Credit: 22,446,667 RAC: 0 |
Forum feels like its working faster now. :) |
Send message Joined: 30 Aug 07 Posts: 2046 Credit: 26,480 RAC: 0 |
Yeah, the server isn't getting nailed nearly has hard as it was :) |
Send message Joined: 22 Nov 07 Posts: 285 Credit: 1,076,786,368 RAC: 0 |
|
Send message Joined: 15 Jul 08 Posts: 383 Credit: 729,293,740 RAC: 0 |
The larger WUs are definitely helping here too. The cache on my X2 is up to 42 minutes and a whopping 84 minutes on the quads. They aren't hammering the server every minute now. Still though, in order to allow BOINC to work correctly and do it's scheduling job, WUs would either have to be larger still or allow enough of them to be able to schedule out at least 6-8 hours. But don't get me wrong, this is an appreciated improvement! |
Send message Joined: 15 Jul 08 Posts: 383 Credit: 729,293,740 RAC: 0 |
It looks like some shorter WUs are working there way back into the system. If the WUs are allowed revert back to the smaller size, the GPU clients will start hammering the server every minute and the crashes will resume. The 4x longer WUs helpped but between the WU size and tiny allowed cache for GPU machines, BOINC is still unable to schedule this project properly unless other GPU projects are suspended. Please consider increasing the WU size even more and increasing the WU queue size, at least for GPUs. Thanks. |
Send message Joined: 6 May 09 Posts: 217 Credit: 6,856,375 RAC: 0 |
Thank you for your concern, Beyond. The work load of the server is one of our major concerns right now, so we are being careful to keep it down. Right now the server is running very smoothly, but we are able to adjust the WUs in short notice, if necessary. Just keep us posted if something gets bad. Thanks! |
Send message Joined: 12 Aug 09 Posts: 172 Credit: 645,240,165 RAC: 0 |
Hi Matthew, the problem is for those of us with high end GPU's we hammer the server every minute or so looking for more work with the 1 stream WU's. Is there any way the 1 stream WU's can be restricted to boxes with CPU only? This would achieve two things: 1. It would keep the low end boxes happy by giving them WU's they can complete in a reasonable time, not feel unwanted by the project, and; 2. It would keep demand down on the server keeping the high end people like myself happy as well. Thanks for the quick responses to everyone's comments. It makes us all feel wanted. |
Send message Joined: 15 Jul 08 Posts: 383 Credit: 729,293,740 RAC: 0 |
Second night in a row that my x2 box with HD4770 GPU ran out of work because of the 39-55 second WUs. Then the server tells the machine not to retry WU fetch for an hour and it sits idle. I'm REALLY REALLY tired of checking it every 10 minutes to see if it has work so it's back to Collatz with that machine. Why the short WUs have been brought back hasn't even been explained to us :-( |
Send message Joined: 12 Nov 07 Posts: 2425 Credit: 524,164 RAC: 0 |
Somebody didn't pay attention and put them at the previous size. Doesn't expecting the unexpected make the unexpected the expected? If it makes sense, DON'T do it. |
Send message Joined: 24 Dec 07 Posts: 1947 Credit: 240,884,648 RAC: 0 |
Second night in a row that my x2 box with HD4770 GPU ran out of work because of the 39-55 second WUs. Then the server tells the machine not to retry WU fetch for an hour and it sits idle. I'm REALLY REALLY tired of checking it every 10 minutes to see if it has work so it's back to Collatz with that machine. Why the short WUs have been brought back hasn't even been explained to us :-( That's what update scripts are for! An update every 10 minutes ensures your cache is never dry for long and gets around the project maintenance backoff and doesn't hit the server too hard. |
Send message Joined: 15 Jul 08 Posts: 383 Credit: 729,293,740 RAC: 0 |
Second night in a row that my x2 box with HD4770 GPU ran out of work because of the 39-55 second WUs. Then the server tells the machine not to retry WU fetch for an hour and it sits idle. I'm REALLY REALLY tired of checking it every 10 minutes to see if it has work so it's back to Collatz with that machine. Why the short WUs have been brought back hasn't even been explained to us :-( Why not just set the back-off at 10 minutes instead of an hour. Having to run scripts to defeat the server settings is a slippery slope and will most likely lead to abuse in the long run. If the server is set up thoughtfully the work outages simply won't happen, much more work will be processed and users won't give up in frustration and leave. |
Send message Joined: 24 Dec 07 Posts: 1947 Credit: 240,884,648 RAC: 0 |
Second night in a row that my x2 box with HD4770 GPU ran out of work because of the 39-55 second WUs. Then the server tells the machine not to retry WU fetch for an hour and it sits idle. I'm REALLY REALLY tired of checking it every 10 minutes to see if it has work so it's back to Collatz with that machine. Why the short WUs have been brought back hasn't even been explained to us :-( I can't disagree with that approach, but what we've seen is a lack of understanding by the project admins on the nuances of setting up BOINC server side. Script use was quite common during the period when work was very difficult to get and I know I had mine set up to update at a period just a little longer than the 'just got work' backoff. With the introduction of the short wu's I wonder if we are going to hit server load issues again? |
Send message Joined: 26 Jan 09 Posts: 589 Credit: 497,834,261 RAC: 0 |
|
Send message Joined: 15 Jul 08 Posts: 383 Credit: 729,293,740 RAC: 0 |
OK, I gave up and installed a script that updates all my MW boxes every 10 minutes. Took a while to get it running the way I wanted and updating all the machines with one script though... |
©2024 Astroinformatics Group