Times (Elapsed / CPU) for BRP5/6/6-Beta on various CPU/GPU combos - DISCUSSION Thread
Language
Copyright © 2024 Einstein@Home. All rights reserved.
Copyright © 2024 Einstein@Home. All rights reserved.
RE: Please feel free to
)
The section with the results
Elapsed Time Statistics CPU time Statistics etc is ok if both GPUs are/were running with the same performance. There is quite a difference between them in my case. Not so much now!
That said, identifying which GPU completed which task (in entirety) is not trivial from the website.
I will probably add additional rows one per GPU .
Another good source for data is the job log /var/lib/boinc-client/job_log_einstein.phys.uwm.edu.txt - which in my case has about has several years worth of data.
I will post when i have about 20 or so completed tasks, but the headline figures are (running three tasks) are GPU0 30Ks -> 20Ks and GPU1 77Ks -> 23Ks.
It should not take too long!
Edit: I would also suggest adding Min and Max to the column headers (either side of the Mean). The variance is not really a useful stat iirc it is only the square of the more commonly used std deviation.
To "Gary and his gang":
)
To "Gary and his gang": thanks for all the work you're putting into profiling the new app! I'm sure it's appreciated by many.
@Gary: is your Host 02 limited to PCIe 2.0 by the mainboard, as is Host 04? With Haswell all CPUs should finally support version 3.0, but e.g. the H81 chip limits them to 2.0.
Is your Host 03 really using PCIe 1.x? A quick look at a price comparison site shows only 2 mainboards with ancient GeForce 7025 / nForce 630a chipsets which do not offer PCIe 2.0 for the current AMD FX CPUs.
@Gavin: if I were you I'd try overclocking the memory on the GTX660Ti. Something between 1.55 and 1.65 GHz should work and yield a nice boost.
MrS
Scanning for our furry friends since Jan 2002
Hi MrS, RE: @Gary:
)
Hi MrS,
Four of the six G3258 Haswell systems I have use Asus H81-M-Plus motherboards and can confirm they are limited to PCI-e v2.x Host no.2 in my first post in the results thread is one of these machines.
I pushed the boat out for the other two 3258's and paired them with Asus Z97-K boards in the hope of unlocking the full potential of these chips and gain PCI-e3... I will work up the stats from one of these machines and post the results later today. You may be very surprised at the difference!
My 660Ti is one of the very early release models and is terrible!! The only reason its back in my mix is because my trusty old GTX580 finally gave up the ghost.
I have tried overclocking the thing in the past, it responded very poorly, but I am happy to try once again.
RE: This is a very
)
Well, I used to had the same PCI Bandwidth utilization for my GTX 770 (~1%) but now somehow it is back to dozen of percents. I'm a bit puzzled why. I'm still crunching beta 1.50 and all my tasks(more then dozen) on this host have approximate the same PCI utilization.
Is it result of deviation?
RE: Please feel free to
)
Something like this....
[pre]
Elapsed Time Statistics CPU time Statistics
---------------------------------- ------------------------------------ Sample
Search Min Mean Max Std Dev Min Mean Max Std Dev Size Notes / Comments
====== ====== ====== ====== ======= ======= ====== ====== ======== ====== ================
GPU0-BRP5 ####### ####### ###### ###### ####### ####### ###### ###### #### *** Add something here if you wish ***
GPU0-BRP6 ####### ####### ###### ###### ####### ####### ###### ###### #### *** Add something here if you wish ***
GPU0-BRP6b ####### ####### ###### ###### ####### ####### ###### ###### #### *** Add something here if you wish ***
GPU1-BRP5 ####### ####### ###### ###### ####### ####### ###### ###### #### *** Add something here if you wish ***
GPU1-BRP6 ####### ####### ###### ###### ####### ####### ###### ###### #### *** Add something here if you wish ***
GPU1-BRP6b ####### ####### ###### ###### ####### ####### ###### ###### #### *** Add something here if you wish ***
[/pre]
RE: Well, I used to had the
)
When are you observing these numbers? At the beginning of WUs (OK) or at some random points (would not sound OK)?
MrS
Scanning for our furry friends since Jan 2002
RE: RE: Well, I used to
)
I am using the nVidia X Server Settings Tool and now see GPU1 up at 50% the other GPU0 4%. Fairly consistently for last 10 minutes.
All 6 tasks are 50-90% complete. Not what I expected.
The PCIe bandwidth usage is
)
The PCIe bandwidth usage is highly data dependent. I hope to make a new beta app version soon (this week) that will further optimize this and will reduce the variance between different WUs, hopefully.
HB
RE: @Gary I will post my
)
I have no problem with your modified format - in fact it's an improvement - much easier to take in the sequence of min -> mean -> max -> SD for both Elapsed and CPU. I'll move your set of results to the RESULTS thread later on.
I don't know what's best to do for multi-GPU setups. Firstly, I don't have any so I haven't put much thought into it. I might be totally wrong but I tend to think that interactions might be such as to make all GPUs sharing the bus have a somewhat 'crippled' performance compared to a single unit. Also this might vary over time so that stats might show worse variation than what is directly attributable to the way the beta app itself now works. I don't want to discourage people from submitting their results - I certainly welcome the submission and sincerely thank you for the effort you have made. I, also, don't expect people to manually separate tasks into different pools based on the different GPUs in the mix. But if people don't do that, the stats results will have something of a question mark hanging over them. I wouldn't reject the submission of unsorted results as long as the fact they were unsorted was clearly stated. It would then be up to the ultimate consumer of the information to use it or otherwise, based on full disclosure. BTW, I very much like all your ***1 , etc, notes scattered throughout - very helpful to the reader.
Once again, sincere thanks for all the effort.
Cheers,
Gary.
RE: The PCIe bandwidth
)
Does it mean that the "empty" WUs will be processed much faster then WUs with possible candidates?