1.28 Gamma-ray pulsar binary search #1 on GPUs (FGRPopencl2-ati)
1 Sep 2021 19:48:48 UTC
Topic 225945
(moderation:
I am now running the new 1.28 GR work units on an RX 570 (Ubuntu 20.04.3).
They are taking around 10 minutes, supported by six cores of a Ryzen 3600.
https://einsteinathome.org/host/12878436/tasks/2/0?sort=desc&order=Sent
This is about the same as the 1.18.
https://einsteinathome.org/content/what-do-you-expect-crunching-einsteinhome#comment-188429
On my 1660ti running Ubuntu
)
On my 1660ti running Ubuntu as well my times have gone from the 750 second range
to the 450 second range
For my Polaris20/i5 test
)
For my Polaris20/i5 test system I noticed the following times (x2, no power/undervolting tool applied):
Mint 20.2 [5.4.0], Boinc 7.16.6:
960 sec: coproc file in original status, so app version 1.18
960 sec: coproc file in trick status, so app version 1.28
Xubuntu 18.04.5 [5.4.0], Boinc 7.9.3:
1015 sec: driver amdgpu-pro 20.40, so app version 1.18
Oddly however, for all 1.28
)
Oddly however, for all 1.28 workunits that I crunched after a pause, I got messages like the following, and they disappeared after uploading without points. (The ones before the pause were all fine.)
2021-09-02 12:08:19.6748 [PID=2287 ] [CRITICAL] [HOST#12900320] [RESULT#1162302446] [WU#571500842] result already over [outcome=1 validate_state=0]: result already reported as success
2021-09-02 12:08:19.6748 [PID=2287 ] [handle] [HOST#12900320] [RESULT#1162312911] [WU#571505665] got result (DB: server_state=5 outcome=1 client_state=5 validate_state=1 delete_state=2)
2021-09-02 12:08:19.6793 [PID=2287 ] [debug] [HOST#12900320] MSG(high) Completed result LATeah4012L00_948.0_0_0.0_1994655_0 refused: result already reported as success
Thereafter, I aborted a few remaining ones. They also disappeared (instead of appearing in the error column) from the dashboard. Odd!
By now I have run many
)
By now I have run many hundreds of 1.28 tasks on four different GPU's residing on three different hosts.
Two 5700s
one 6800 XT
one 6800
While I have not seen the failures that others have reported, I also have seen extremely little performance benefit. Tentatively, I think I see about 1% performance benefit in two cases where the GPU was running tasks at 2X multiplicity while the other cases where it was running 3X or 4X have no discernible benefit at all.
Thanks arch. could you run a
)
Thanks arch. could you run a few tests on both a Navi and Big Navi system doing 1x w/1.22 and 1x w/1.28? Thats where I observed the reported 20% increase in pre-release, with 1x. I do think the app still needs some tuning though since it's not performing the same as pre-release code. so something else might need to be added.
The failures that most people have seen, seem to be driver related.
Windows drivers seem to not have proper OpenCL support for Vega or older (even though they report they do), so they produce errors. I have not seen a single host that was Windows based with Vega or Polaris gen cards with a successful run. I took my Polaris RX570 GPU and entire platform that worked on Linux, and it would not work on Windows.
I do not believe it to be a problem with the windows application itself since you (arch) have been successful in running Windows drivers with Navi and Big Navi. Wedge had issues with one of his hosts on Linux with ROCm drivers and a Vega GPU, but I believe his issue to be hardware incompatabiliity (very old CPU, which seems to be important to ROCm drivers) and his failure mode was different than everyone else's (no error, tasks just didnt run, no progress, no GPU utilization).
Linux hosts with ROCm drivers have been shown to work on several hosts with the new app with Polaris cards and Vega cards (no change in speed, or slightly slower) and with Navi cards as well (small speedup).
_________________________________________________________________________
Linux Mint 19.3 Xfce,
)
Linux Mint 19.3 Xfce, amdgpu-pro 20.30, Radeon VII, x 3:
1.18: 472 s
1.28: 520 s
Ian&Steve C. wrote:run a few
)
I wondered if you perhaps had run the tests at 1X. I don't think any serious Einstein Navi user runs FGRP that way, so if it turns out to be true that the benefit is strongly multiplicity dependent, that blows a big hole in the case for any AMD benefit worth having.
I've started work toward generating 1X comparisons, first on a 5700 system, then on a 6800 system. I should have at least one comparison before the end of the day.
I, in turn, have a request of you. Please don't refer to me by a diminutive. I'm fine with you calling me archae86 or Peter, but don't wish to be called arch or Pete. Thank you.
Tom's gains on his 5700
)
Tom's gains on his 5700 continued to show improvment on 2x also. it wasnt just on 1x.
edit:
to be more accurate, Here were the results we observed with Tom's 5700
So there was about 8-10% advantage for 2x operations with test code, but only ~2-3% advantage with the 1.28 app.
I was just looking for a more apples to apples comparison to be sure that there are still only marginal gains when running 1x.
the new 1.28 amd app more than likely is still missing some of the key optimizations petri made in our test code, and Bernd has informed me that he will be away for most of September, so it's unlikely to see any update until next month.
_________________________________________________________________________
I completely missed that this
)
I completely missed that this was an AMD/ATI thread; sorry about that. I was so smitten with "Petri33 was here" ... well,
#idiot
catavalon21 wrote: I
)
I am perfectly happy to see Nvidia results. I just happen to have my RX 570 on Einstein most of the time, since that is what it does best. Why don't you start a separate thread for Nvidia, to keep them straight?