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 : Number crunching : Geminids Shower Challenge (Message 152604)
Posted 41 days ago by xii5ku
Well, both the <app> and the <app_version> variants work for the basic purpose of specifying just the CPU count and GPU count per task. One or the other (or both) need to be used if more than that is desired. E.g. <app> for concurrency, <app_version> for command line arguments.
2) Message boards : Number crunching : Geminids Shower Challenge (Message 152602)
Posted 41 days ago by xii5ku
@bcavnaugh,
the syntax for an app_config which configures both the CPU version and the GPU version should look like this:
<app_config> <app_version> <app_name>genefer</app_name> <plan_class>cpuGFN21</plan_class> <cmdline>-nt 8</cmdline> <avg_ncpus>8</avg_ncpus> </app_version> <app_version> <app_name>genefer</app_name> <plan_class>OCLcudaGFN</plan_class> <avg_ncpus>0.658690</avg_ncpus> <ngpus>1</ngpus> </app_version> </app_config>


However, most people don't need an app_version block for the GPU version.

Additional app names and plan_class identifiers can be looked up at apps.php. E.g., Genefer 22 on an AMD card would have genefer_wr as app_name and openclatiGFN22 as plan_class.
3) Message boards : Number crunching : Martin Gardner's Birthday Challenge (Message 151975)
Posted 87 days ago by xii5ku
I stumbled over this previous post of mine:
Two months ago in the Blue Moon challenge thread xii5ku wrote:

arakelov wrote:
Now that the challenge is about to finish and I'll change preferences to download tasks for other subprojects, what do you recommend? Should I let SMT inactive? Are there other applications that prefer SMT on?

The currently active PrimeGrid subprojects which are worthwhile to run on the CPU (instead of the GPU) all look to me as if they are similar in the regard that some might have a very small throughput benefit from SMT, while in other cases SMT decreases throughput. However, I am not sure about the Ryzen 7 3700X in particular. However, even in cases when the currently active PrimeGrid subprojects benefit from SMT with respect to throughput, there will always be a regression in electric energy spent on the same amount of work (expressed in Joule per credit, or in Watts per points-per-day). Hence it is generally preferable not to use SMT when running PrimeGrid CPU-only applications, in my experience.

Before the Gardner's B-day challenge, I had been measuring performance and performance-per-Watt on my computers with a fixed test WU again, for GCW-LLR. Results on Broadwell-EP confirmed what I said above about lower perf/W when HT is used. But to my surprise, an Epyc Rome behaved differently: Of the several configurations in which I used SMT, the best ones matched or very slightly exceeded the perf/W of the best tested non-SMT configs.

I have to say though that while my error of measurement for performance is quite low, there is a bit more error involved in my power measurements. They consist of me periodically looking at the wall power meter. In the mentioned case, it would have been better to measure the actual task energy (i.e. Wh, not W). This would have required either a logging power meter or a lot more attention during the tests, of which I had neither.
4) Message boards : Number crunching : Martin Gardner's Birthday Challenge (Message 151906)
Posted 92 days ago by xii5ku
On Oct 16 Grebuloner wrote:
What is the current FFT range for tasks?

I looked up the sizes as chosen on Haswell for all "b" and "min n in progress"/ "max n in progress"/ "max n loaded" on October 12:

13 ___ 4234840/ 4330158/ 4343034 __ 1920K
29 ___ 3259942/ 3298500/ 3308112 __ 2016K
47 ___ 2850640/ 2884836/ 2893336 __ 2M/ 2304K/ 2304K
49 ___ 2817958/ 2853898/ 2862348 __ 1792K
55 ___ 2707792/ 2771662/ 2779828 __ 2M/ 2304K/ 2304K
69 ___ 2581852/ 2623208/ 2630906 __ 2304K
101 __ 2379576/ 2406490/ 2413702 __ 2304K
109 __ 2323728/ 2367570/ 2374138 __ 2304K
121 __ 2290840/ 2316018/ 2322678 __ 1920K

The same, again on Haswell, but with the b and n range from right now:

13 ___ 4281736/ 4576236/ 4601952 __ 1920K/ 2016K/ 2016K
29 ___ 3270760/ 3485662/ 3505386 __ 2016K/ 2304K/ 2304K
47 ___ 2850640/ 3048694/ 3065802 __ 2M/ 2304K/ 2304K
49 ___ 2830108/ 3015928/ 3032950 __ 1792K/ 1920K/ 1920K
55 ___ 2750254/ 2929120/ 2945490 __ 2304K
69 ___ 2593972/ 2772220/ 2787712 __ 2304K
101 __ 2386636/ 2543368/ 2557572 __ 2304K/ 2400K/ 2400K
109 __ 2352808/ 2501962/ 2516052 __ 2304K/ 2400K/ 2400K
121 __ 2293086/ 2447536/ 2461276 __ 1920K/ 2016K/ 2016K

5) Message boards : Number crunching : Once in a Blue Moon Challenge (Message 151236)
Posted 152 days ago by xii5ku
Bur wrote:
I would assume the scheduler will shuffle the threads from core to core

No, why would it? At least on Linux, the scheduler lets a software thread remain on the same logical CPU as long as feasible, because that's a more efficient use of the caches than kicking a thread unnecessarily around different logical CPUs. Furthermore, the scheduler is aware of HT/SMT hardware thread siblings among the logical CPUs and attempts to spread the overall system load across all physical cores. (This is from memory and aligns with what I observe, alas I don't have links to sources.)
6) Message boards : Number crunching : Once in a Blue Moon Challenge (Message 151233)
Posted 152 days ago by xii5ku
arakelov wrote:
For this challenge I turned SMT off in my BIOS following your suggestions (and it really made a difference!)

There is an alternative: Leave SMT on but let BOINC use only half of the logical CPUs. You have got a Linux computer; the process scheduler of the Linux kernel pursues a policy of spreading the total workload across all physical cores, generally.


arakelov wrote:
Now that the challenge is about to finish and I'll change preferences to download tasks for other subprojects, what do you recommend? Should I let SMT inactive? Are there other applications that prefer SMT on?

The currently active PrimeGrid subprojects which are worthwhile to run on the CPU (instead of the GPU) all look to me as if they are similar in the regard that some might have a very small throughput benefit from SMT, while in other cases SMT decreases throughput. However, I am not sure about the Ryzen 7 3700X in particular. However, even in cases when the currently active PrimeGrid subprojects benefit from SMT with respect to throughput, there will always be a regression in electric energy spent on the same amount of work (expressed in Joule per credit, or in Watts per points-per-day). Hence it is generally preferable not to use SMT when running PrimeGrid CPU-only applications, in my experience.
7) Message boards : Number crunching : Sier"pi"nski's Birthday Challenge (Message 149338)
Posted 316 days ago by xii5ku
xii5ku wrote:
In order to find out whether or not to use SMT with a given CPU, the SMT-on and SMT-off configurations need to be tested with the very same workunit.

If you can't (or don't want to) perform multiple tests with the same WU, then you (1.) need more tests, (2.) need to normalize the test results to the granted credit.

Re, normalization to credit per result: I just looked at some llrSOB results of a Windows computer which is not my own. The PPD which I can calculate from that still has a high variability. Still too high to properly determine what actual difference SMT-on and SMT-off really make.

However, given the cache size and RAM performance requirements of llrSOB, it is pretty safe to say that SMT does not help at with llrSOB in particular.

(Personally, I am not guessing how to configure my computers optimally. I rely on reproducible and precise tests instead. I currently have two types of computers myself, and took all measurements which I need on one of these computer types within just a few hours last evening, using the same WU in all measurements.)
8) Message boards : Number crunching : Sier"pi"nski's Birthday Challenge (Message 149319)
Posted 317 days ago by xii5ku
Bur wrote:
SMT [...] should always be slower.

With some LLR based sub-projects (maybe not llrSOB, I don't know) it gives slightly higher throughput on some processors.

Though at the few occasions at which I measured not only throughput but also power consumption, SMT gave lower throughput/Watt i.e. higher energy use per task.

Bur wrote:
the windows scheduler

I am lazy and use an OS with a sensible scheduler instead.
9) Message boards : Number crunching : Sier"pi"nski's Birthday Challenge (Message 149310)
Posted 317 days ago by xii5ku
Chooka wrote:
Running all 8 threads, it takes 5.48 days to complete 8 tasks.

Correction: To complete one task.


Chooka wrote:
I'm struggling to work out if it's better to run 8T or 4C based on your figures

My guess is that each row represents one test, and that a different, random workunit was used in each of these tests. (The numbers which were tested could have been put into this report to make it more useful. Or the credit granted for the result. Or even better: Both.)

In order to find out whether or not to use SMT with a given CPU, the SMT-on and SMT-off configurations need to be tested with the very same workunit.

If you can't (or don't want to) perform multiple tests with the same WU, then you (1.) need more tests, (2.) need to normalize the test results to the granted credit.
10) Message boards : Number crunching : Sier"pi"nski's Birthday Challenge (Message 149297)
Posted 317 days ago by xii5ku
Michael Goetz wrote:
A head's up to anyone who was planning on using cloud computing resources from Digital Ocean, [...] basic droplets no longer support AVX512, and we have to use the premium droplets.

Before you concern yourselves with per-core performance of your machines (virtual or not), you need to know how much last-level cache the application instance can grab. And if this does not amount to "a lot", then you would want to know how much main memory bandwidth the application instance can have for itself.


Allen Paschke wrote:
Each of my 3 computers, each running a SOB task with all 4 threads, will complete one SOB task during the challenge. It will take each computer 6 - 7 days to complete a task.

Check whether these computers have single-channel or dual-channel memory. If the former, consider to purchase another SO-DIMM. Maybe the KabyLake-U can then make it in 5 days (if the thermal budget allows it); I don't know about Sandy and Ivy Bridge-M.


Next 10 posts
[Return to PrimeGrid main page]
DNS Powered by DNSEXIT.COM
Copyright © 2005 - 2022 Rytis Slatkevičius (contact) and PrimeGrid community. Server load 2.14, 2.46, 2.60
Generated 22 Jan 2022 | 4:23:50 UTC