PrimeGrid
Please visit donation page to help the project cover running costs for this month

Toggle Menu

Join PrimeGrid

Returning Participants

Community

Leader Boards

Results

Other

drummers-lowrise
1) Message boards : Generalized Fermat Prime Search : GFN-1x Small Primes search starts 2021-01-21 (Message 150254)
Posted 51 days ago by Peter Hucker
With other projects (Einstein, Milkyway, Collatz), I've been running 2 or more tasks at once on the GPU to give extra throughput when waiting for CPU assistance. Should I be doing that on this project and on the main Primegrid?

This project utilizes GPU very well, running 2 tasks may give better throughput only on fastest GPUs (top 20xx or 30xx models). A generic rule is to check GPU usage in your real setup first (Windows: use freeware GPU-Z utility, Linux: nvidia-smi installed together with drivers). If GPU usage is above 95%, running 2 or more tasks will not give noticeable effect.

Yes that's what I normally do, if the percentage is 85% or less, I add another task. My fastest card is only an R9 280X, but I have slow (or otherwise busy) CPUs.

The reason I asked is because I was wondering if a similar thing happens with Prime on GPU as on CPU, overloading the cache, etc., and adding more tasks would result in less overall throughput.
2) Message boards : Generalized Fermat Prime Search : GFN-14/13/... MEGA Prime Search on GFN Server (Message 150246)
Posted 52 days ago by Peter Hucker
I would guess that the i5-8600k and Xeon X5650 would do best running 3 tasks with 2 cores each, and that the N3700, Q8400 and i3-350M would do best running a single task with all cores, due to their smaller cache sizes.

You can turn hyperthreading off in the BIOS if you want, or you can tell BOINC to only use 50% of the CPUs if you leave hyperthreading on. For the dual Xeon system, it will also be important to use a program like Affinity Watcher or Process Lasso to make sure that the tasks are staying on the same CPU.

These estimates might apply to the MEGA project, but LLR2 Testing is well above that point right now in terms of test size, so they might be a bit off, but they'd probably be close.

Enjoy :)

Thanks very much. I run other projects, but I've started concentrating on one at a time (well one for CPU and another for GPU sometimes). Some projects don't care if HT is on, most go faster with HT, but Prime goes slower. So I'll set Boinc to 100% or 50% accordingly when I change project.

I'll take your word for it on the number of cores per task for each computer, as I don't feel like running a load of tests all day!
3) Message boards : Generalized Fermat Prime Search : GFN-1x Small Primes search starts 2021-01-21 (Message 150245)
Posted 52 days ago by Peter Hucker
With other projects (Einstein, Milkyway, Collatz), I've been running 2 or more tasks at once on the GPU to give extra throughput when waiting for CPU assistance. Should I be doing that on this project and on the main Primegrid?
4) Message boards : Generalized Fermat Prime Search : GFN-14/13/... MEGA Prime Search on GFN Server (Message 150243)
Posted 52 days ago by Peter Hucker
Thanks, shall I use 3 threads as per the MEGA?

It depends on CPU. This is a Proth test and it scales well, so Intel is OK with 4 cores. I think that there is no need to go higher - when we finish k=7, tasks will be shorter and multithreading will be less efficient. On Ryzen, number of cores must match chip structure to keep task within single CPU block.

And is this right?

Yes, it's fine. Everything is same for all apps, just change application name.

I can't find Ryzen chip structure anywhere, I've seen it before somewhere. Google is not working! Any mention of "block" gives me water cooling blocks. And structure just gives me general specs and gaming benchmarks. It's a Ryzen 3900XT if anyone knows how they look inside. All my other machines are Intel, from an i5 8600K to an ancient DDR2 thing.


For the 3900XT, your best throughput right now will still be single-threaded. The second best would probably be 3 threads per task; but only if you are using some software like Affinity Watcher or Process Lasso to keep the tasks on adjacent cores in the CPU. The 3900XT can be thought of as 4 separate CPUs with 3 cores each. They can communicate with each other, but relatively slowly, so if you keep the tasks in each mini-CPU (CCX is the proper term in this case), you will see the best performance for multi-threading LLR tasks, as long as you stay under the 16MB cache limit.

Ultimately, for the 3900XT, for the remainder of the 100,000,000+ tasks on LLR2 Testing on Stream's server; single-threaded is your best bet.

Thanks. Sorry to ask, but can you give me an estimation of what's best for my other machines? They're all Intels. i5 8600K (6 core), Xeon X5650 (6 core + HT) x 2 CPUs per machine, and 3 older ones: N3700 (4 core), Q8400 (4 core), i3 350M (2 core + HT).

Also I'm assuming it's best to turn the HT off?

Do these estimates apply to the MEGA project?
5) Message boards : Generalized Fermat Prime Search : GFN-14/13/... MEGA Prime Search on GFN Server (Message 150237)
Posted 52 days ago by Peter Hucker
Thanks, shall I use 3 threads as per the MEGA?

It depends on CPU. This is a Proth test and it scales well, so Intel is OK with 4 cores. I think that there is no need to go higher - when we finish k=7, tasks will be shorter and multithreading will be less efficient. On Ryzen, number of cores must match chip structure to keep task within single CPU block.

And is this right?

Yes, it's fine. Everything is same for all apps, just change application name.

I can't find Ryzen chip structure anywhere, I've seen it before somewhere. Google is not working! Any mention of "block" gives me water cooling blocks. And structure just gives me general specs and gaming benchmarks. It's a Ryzen 3900XT if anyone knows how they look inside. All my other machines are Intel, from an i5 8600K to an ancient DDR2 thing.
6) Message boards : Generalized Fermat Prime Search : GFN-14/13/... MEGA Prime Search on GFN Server (Message 150232)
Posted 52 days ago by Peter Hucker
Does the "LLR2 testing" subproject do multithreading aswell?

Yes, and it's currently recommended (at least until we drop down below n=2M). Application name for app_config.xml is llr2

Thanks, shall I use 3 threads as per the MEGA?

And is this right?


<app_config> <app_version> <app_name>llr2</app_name> <cmdline>-t 3</cmdline> <avg_ncpus>3</avg_ncpus> </app_version> </app_config>
7) Message boards : Generalized Fermat Prime Search : GFN-14/13/... MEGA Prime Search on GFN Server (Message 150225)
Posted 53 days ago by Peter Hucker
Does the "LLR2 testing" subproject do multithreading aswell?
8) Message boards : Number crunching : Boinc refuses to get new work from Primegrid, it says "don't need". (Message 150113)
Posted 65 days ago by Peter Hucker
Maybe you set "Max# of primegrid tasks" in the primegrid prefs page to 1?

No. It's on unlimted.


3128 PrimeGrid 16-04-2021 11:38 AM Not requesting tasks: don't need (CPU: ; AMD/ATI GPU: )


When you see "Not requesting tasks" here, the problem is completely on client side, because client didn't asked for any work from server. In complex setups with app_config(s), finding an exact reason may be difficult (like in this case, where Boinc said "don't need" but didn't said WHY it didn't need any tasks - sometimes it explains the reason after "CPU:" and "GPU:" words). A common rule is to double-check all config overrides in app_config.xml or temporary disable them (full restart of Boinc service is required, just reboot your PC if not sure).

As I said in the update:

The problem is I had excluded two of the many subprojects of Primegrid from one of the cards (the slower one) in cc config. Because the one I had running was one of those, presumably it didn't want to get more incase they were also unsuitable for the other card.

So Boinc being a little daft in two ways. Firstly, it should have seen the card is supposed to run two at once. Secondly, there seems to be no communication with the server about excluded apps. So if I tell it to not run subproject X on GPU 1, then when GPU 1 runs out of work, it will just request "work" from the server, without telling it which subprojects are required. I've seen a similar thing when I limit tasks from LHC or Kryptos using virtualbox to 4 at a time in app config (since more virtualboxes slows the machine terribly), it continues to download more and more of them, then tries another project. I ended up with a week's work of Kryptos waiting to do 4 at a time with the other 20 threads doing other projects, when my buffer was supposed to be 3 hours.

Anyway I've worked around the problem, I've banned genefer extreme from this machine on the server end. The problem I had was if genefer extreme runs on one card, then is switched to the other card after rebooting or an exclusive application, I get a wrong result (it fails verification). The other subprojects seem ok using more than one card.
9) Message boards : Number crunching : Boinc refuses to get new work from Primegrid, it says "don't need". (Message 150110)
Posted 65 days ago by Peter Hucker
Suddenly one of my machines won't get Primegrid GPU work. It's the only project with GPU tasks available that isn't set to "no new work". I've tried manually updating. The machine has two graphics cards, and one task currently running, and none queued. I've told it the tasks use 0.5 GPUs in app config, so it needs 4. Its (badly) estimated remaining time is 9 days (it will actually take about 3), with a deadline of 14.5 days, and it's not in panic mode (doesn't say "running high priority"). I've tried restarting Boinc. I get this (let me know if some other debug options would give the answer):

Ryzen 3124 PrimeGrid 16-04-2021 11:38 AM update requested by user 3125 PrimeGrid 16-04-2021 11:38 AM [sched_op] sched RPC pending: Requested by user 3126 PrimeGrid 16-04-2021 11:38 AM [sched_op] Starting scheduler request 3127 PrimeGrid 16-04-2021 11:38 AM Sending scheduler request: Requested by user. 3128 PrimeGrid 16-04-2021 11:38 AM Not requesting tasks: don't need (CPU: ; AMD/ATI GPU: ) 3129 PrimeGrid 16-04-2021 11:38 AM [sched_op] CPU work request: 0.00 seconds; 0.00 devices 3130 PrimeGrid 16-04-2021 11:38 AM [sched_op] AMD/ATI GPU work request: 0.00 seconds; 0.00 devices 3131 PrimeGrid 16-04-2021 11:38 AM Scheduler request completed 3132 PrimeGrid 16-04-2021 11:38 AM [sched_op] Server version 713 3133 PrimeGrid 16-04-2021 11:38 AM Project requested delay of 7 seconds 3134 PrimeGrid 16-04-2021 11:38 AM [sched_op] Deferring communication for 00:00:07 3135 PrimeGrid 16-04-2021 11:38 AM [sched_op] Reason: requested by project


Update. The problem is I had excluded two of the many subprojects of Primegrid from one of the cards (the slower one) in cc config. Because the one I had running was one of those, presumably it didn't want to get more incase they were also unsuitable for the other card.

But.... The card that was permitted to run any only had one task running, and was set to run two at once, so why didn't it get another?
10) Message boards : Number crunching : Too much work when joining (Message 149969)
Posted 71 days ago by Peter Hucker
That would mean I should be fine using some of my threads for Primegrid and some for a different type of program. Then again, would the OS know what to pair with what? Or does the CPU sort that stuff out?

That's up to the OS to figure out scheduling. As for knowing what to do with it... linux tends to do better in that regard.

I'm sure I've read about modern CPUs doing something like this themselves, could have been something similar.


Next 10 posts
[Return to PrimeGrid main page]
DNS Powered by DNSEXIT.COM
Copyright © 2005 - 2021 Rytis Slatkevičius (contact) and PrimeGrid community. Server load 1.84, 2.19, 1.99
Generated 20 Jun 2021 | 18:02:17 UTC