Gamma-Ray Pulsar search (FGRPB1) CPU app version 1.08
Bernd Machenschalk
Moderator
Administrator
Joined: 15 Oct 04
Posts: 4312
Credit: 250362588
RAC: 35541
26 Jul 2017 19:41:56 UTC
Topic 208951
(moderation:
Since we had to retract app version 1.06/7, I just issued version 1.08.
This version comes with an own fftw-wisdom binary (should be in the project directory once you downloaded the app, named fftwf-wisdom_FGRPB1_1.08_*) to create "wisdom". Run this with arguments "-o FGRPB1wisdom.dat rib67108864" and put the file "FGRPB1wisdom.dat" into the project directory.
This version is published as "Beta" first, so we can check validation.
n/a
A couple of questions. 1.
)
A couple of questions.
1. Should all other activities on the host be stopped when running the wisdom binary?
2. Does it make any difference (apart from time to run) if the host is also crunching?
3. Can a wisdom file generated on one host be used on another with similar hardware?
4. If so, how similar does 'similar' need to be?
5. For example would same CPU and motherboard but different GPU be OK?
I've created the file on one host (took around 30 mins) with BOINC shut down. After restarting BOINC, I promoted one of the beta tasks and it has now completed. It took about 3% longer to crunch than the average of the 15 previous tasks most recently crunched on that machine. Is a longer elapsed time likely to be experienced by some types of machines? Are there any known hardware types that are likely to benefit?
Here is a link https://einsteinathome.org/host/431681/tasks/0/32 to the FGRPB1 tasks on that machine. The beta task is currently at the bottom of the page. It has already validated.
(Seems like [url] ... [/url] doesn't work anymore)
Cheers,
Gary.
Hi Gary! - Ideally
)
Hi Gary!
- Ideally everything else should be stopped while generating wisdom. The impact of other processes running in parallel depends largely on quite a number of parameters and isn't generally predictable.
- AFAIK wisdom only depends on the CPU; there's no dependency on memory, chipset or whatever else.
- From the very few samples so far I think this app version with FFTW 3.3.6 (without wisdom) might be slightly slower than the 1.05, but I have not enough data yet to verify that in general.
- Your task shows "% fft_size: 67108864" (which means that the wisdom should apply) and "Successfully read wisdom from file ..." which means just that (wisdom could be read). I really don't know why this task ran longer rather than shorter.
BM
(Seems like [url] ... [/url]
)
Yep. BBCode is currently broken, devs are on it (https://dev.gridrepublic.org/browse/DBOINCP-396).
BM
Hi Bernd, Thanks very much
)
Hi Bernd,
Thanks very much for the responses.
Cheers,
Gary.
- From the very few samples
)
There aren't much results/d from 1.08 yet, but it looks like on avarage 1.08 is less than 1% slower than 1.05. More importantly, though, all results of 1.08 that have been compared to that of 1.05 are found valid, i.e. the problem we had with 1.06/7 seems to be gone, which is what worried me most. So now we can actually start experimenting with the effect of wisdom.
BM
Thanks BerndA day or so ago
)
Thanks Bernd
A day or so ago i changed 12242223
to run 1.08 tasks only and completion times settled at 38K seconds, about 1% slower than earlier versions.
Last night I added the wisdom file (as below), and the times have increased, probably up to 80K seconds - there are no tasks completed yet fully running the new wisdom.
Tasks for 12242223
I'll leave it running for today. If you want a copy of the wisdom file posted / PMd let me know.
edit: ok, so results completed at 78-79K seconds. I decided to rerun the wisdom this time while the system was under full load. These will look likely ~50K seconds.
Bernd Machenschalk wrote:I
)
I wondered if it might have been due to the age of the CPU. It was a G540 - a Sandy Bridge Celeron dual core that was first released in 2011. As a result of the slower time, I put it back on non-beta tasks.
I decided to try a rather more modern CPU - a Haswell G3260 Pentium dual core first released in 2015. It's the most modern architecture I have.
1.05 CPU tasks average around 21,900 secs mostly with quite a small variation on either side (average of last 23 tasks). The first 1.08 task took 21,090 secs, which is about 3.7% speedup this time. Of course it's only one result but at least it's heading in the right direction this time.
The wisdom file for this one was actually generated whilst the machine was crunching (CPU + RX460 GPU (x2)). I thought I had stopped BOINC but it turned out I hadn't. Too bad, I decided to use it anyway.
I've installed that wisdom file on 3 other hosts with G3260 CPUs and RX460 GPUs as well, and set them up for beta tasks. It will be interesting to see if all 4 hosts get a speedup compared to the standard app, once all have crunched a reasonable number of tasks. At some point I'll generate fresh wisdom on a host with BOINC really shut down to see if that makes any further difference.
Cheers,
Gary.
AgentB wrote:Tasks for
)
Spoke too soon, as the old wisdom tasks completed, they started running down to 33-34K sec so that is at least a 10% improvement on the "no wisdom" so that's good news. It may fall a little further
edit+: it has to 30.5K sec
Might try running again to see if it gets wiser later on...
edit: the logic being the wisdom tunes the FFT based on the running environment.
So the first wisdom was created "no tasks running", the second (currently running) was created when "all tasks running - but slowly", the next one will be created "all tasks running - faster than normal."
I think i'll adopt the repeat wisdom until no further improvement, and see where that goes.
I've run these now with and
)
I've run these now with and without wisdom on three hosts. Somewhat confusing results. Two of the hosts have basically the same processor and speed (Xeon X5660). Wisdom didn't have a notable effect on the first one (~23600 sec), but the other has been slower without wisdom (went from 24700 to 28400 sec).
Third host has old Intel Q9550. Based on minimal test it was clearly faster with wisdom (20800 vs 15500 sec).
I've run a couple of 1.08
)
I've run a couple of 1.08 tasks with wisdom on a Skylake machine. It was some 25 percent faster - thanks to the Einstein engine room!
However the very first of those tasks somehow missed to read the wisdom file although it had been provided in the same directory - maybe that task had run a second or two right after downloading, before creation of wisdom began?