| Index | Recent Threads | Unanswered Threads | Who's Active | Guidelines | Search |
 |
World Community Grid Forums
Category: Active Research
Forum: Africa Rainfall Project
Thread: 3.5 day deadline sent to unsuitable machine |
| No member browsing this thread |
|
Thread Status: Active Total posts in this thread: 25
|
|
| Author |
|
|
Link64
Advanced Cruncher Joined: Feb 19, 2021 Post Count: 129 Status: Offline Project Badges: 
|
Well, you might still be loosing lots of computing time if you need to shut down every day. My computer checkpoints about every 4 hours on ARP, so if I was running ARP on 4 cores and shuting down every day like you (I have only 2 cores, but that's not the point), I'd be loosing up to 16 hours of computing time. Every day.
----------------------------------------Perhaps in this case you should indeed switch to MMC and/or OPN, both ARP and MIP are not suitable for computers on which the WUs can't run from the beginning till the end without interruptions.  [Edit 1 times, last edit by Link64 at Aug 5, 2021 8:39:40 AM] |
||
|
|
kwolff88
Cruncher Joined: Dec 31, 2004 Post Count: 19 Status: Offline Project Badges: 
|
Already tested. Using half of the core not making big difference on this machine (no logic core) Really? I have a 24-core Threadripper and have found that the speed of AFP changes almost linearly with cores, to the point where tasks run twice as fast with 8 running as they do with 22 running. When I had 18 running, it was somewhere in between. I would try running 1 at a time, and see how long it takes it to run; then compare with 2, then 3. If you can get a work unit done in under 3.5 days at 1 unit at a time, that's what you should do. Doesn't give you a lot of runtime/credits, but as others have pointed out, AFP is a very intense project and is best run on more powerful computers, and particularly ones that can be left on 24 hours a day. [Edit 1 times, last edit by kwolff88 at May 28, 2021 1:26:25 PM] |
||
|
|
bidon
Cruncher Joined: Dec 9, 2005 Post Count: 26 Status: Offline Project Badges:
|
The ram bandwidth is the "Achilles' heels" of the Threadrippers, more than others CPU which are often "RAM limited".
And it's seems that some version of Windows had issues with the NUMA of those CPU so they are faster when using Linux. The ARP use larger amount of RAM than other projects so we can suspect it need more RAM bandwidth than others projects. |
||
|
|
Mike.Gibson
Ace Cruncher England Joined: Aug 23, 2007 Post Count: 12364 Status: Offline Project Badges: 
|
I have 20 GB RAM so I don't notice any slowing until I get to 4 of 8 threads (4 cores). I am currently running 3 ARP, 3 MCM and 2 MIP.
----------------------------------------I have a wired network here. I am due to have fibre optic broadband installed on 11 June so will see how much effect that has. Mike [Edit 1 times, last edit by Mike.Gibson at May 29, 2021 12:43:53 PM] |
||
|
|
bidon
Cruncher Joined: Dec 9, 2005 Post Count: 26 Status: Offline Project Badges:
|
Mike, i don't talk about the memory amount but the memory bandwidth.
The RAM is very slow. One core can saturate an dual channel DDR4 controller. That's the reason why there are cache level. Some task does many calculation on a very few data amount. The ram bottleneck is not an issue. You can exploit the threadripper without problem. Some task need more data than the cache L1/L2/L3 can store and few calculation. Core are starved and waiting for data and become slower. There is an good example here : https://www.pugetsystems.com/labs/articles/Ad...lti-Core-Performance-625/ You can see that some "simple" effect can't exploit all cores and "complex" effect can exploit all the cores. There is others review which showing the ram bottleneck issue. |
||
|
|
|