Linux CUDA 'Special' App finally available, featuring Low CPU use

Message boards : Number crunching : Linux CUDA 'Special' App finally available, featuring Low CPU use
Message board moderation

To post messages, you must log in.

Previous · 1 . . . 71 · 72 · 73 · 74 · 75 · 76 · 77 . . . 83 · Next

AuthorMessage
TBar
Volunteer tester

Send message
Joined: 22 May 99
Posts: 5204
Credit: 840,779,836
RAC: 2,768
United States
Message 1902658 - Posted: 24 Nov 2017, 8:13:05 UTC

Same on this one, https://setiathome.berkeley.edu/workunit.php?wuid=2756289395

setiathome v8 enhanced x41p_zi3v, Cuda 8.00 special
Spike count: 0
Autocorr count: 0
Pulse count: 2
Triplet count: 5
Gaussian count: 0

setiathome v8 enhanced x41p_zi3v, Cuda 9.00 special
Spike count: 0
Autocorr count: 0
Pulse count: 3
Triplet count: 5
Gaussian count: 0

SSE4.1xjf OS X 64bit Build 3344
Spike count: 0
Autocorr count: 0
Pulse count: 3
Triplet count: 5
Gaussian count: 0
ID: 1902658 · Report as offensive
Profile Raistmer
Volunteer developer
Volunteer tester
Avatar

Send message
Joined: 16 Jun 01
Posts: 6325
Credit: 106,370,077
RAC: 121
Russia
Message 1902673 - Posted: 24 Nov 2017, 9:28:22 UTC - in response to Message 1902658.  
Last modified: 24 Nov 2017, 9:29:56 UTC

Same on this one, https://setiathome.berkeley.edu/workunit.php?wuid=2756289395

setiathome v8 enhanced x41p_zi3v, Cuda 8.00 special
Spike count: 0
Autocorr count: 0
Pulse count: 2
Triplet count: 5
Gaussian count: 0

setiathome v8 enhanced x41p_zi3v, Cuda 9.00 special
Spike count: 0
Autocorr count: 0
Pulse count: 3
Triplet count: 5
Gaussian count: 0

SSE4.1xjf OS X 64bit Build 3344
Spike count: 0
Autocorr count: 0
Pulse count: 3
Triplet count: 5
Gaussian count: 0


Not quite the same cause it's non-overflow.
Pulse missing for CUDA8. It's same is reported before indeed.

And this makes 4th issue in list (only for particular build it seems).
SETI apps news
We're not gonna fight them. We're gonna transcend them.
ID: 1902673 · Report as offensive
JohnDK Crowdfunding Project Donor*Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 28 May 00
Posts: 1222
Credit: 451,243,443
RAC: 1,127
Denmark
Message 1902688 - Posted: 24 Nov 2017, 11:31:55 UTC - in response to Message 1902625.  

You may want to move to a newer version of Mint, say 18.2. Some Machines and OSes don't respond well to the CUDA 9.0 drivers. I still can't use the CUDA 9 drivers with my older machines using Kernel 4.10. They work fine with Kernels below 4.10 with the CUDA 9 drivers. Your case may be the opposite, you may need to use a newer Kernel with your newer CPU.

Even if there might be problems with cuda9 and mint18.1, shouldn't BOINC still detect the GPUs?
ID: 1902688 · Report as offensive
rob smith Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer moderator
Volunteer tester

Send message
Joined: 7 Mar 03
Posts: 22158
Credit: 416,307,556
RAC: 380
United Kingdom
Message 1902690 - Posted: 24 Nov 2017, 11:43:03 UTC

It should, provided the version of the drivers have all the appropriate computational stuff in them.
Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?
ID: 1902690 · Report as offensive
JohnDK Crowdfunding Project Donor*Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 28 May 00
Posts: 1222
Credit: 451,243,443
RAC: 1,127
Denmark
Message 1902693 - Posted: 24 Nov 2017, 11:49:32 UTC

It would be nice to know if anybody using Mint 18.2 and repo 384.90 drives runs OK... I hate updating version/kernel since I have had trouble several times in the past with Linux not booting afterwards.
ID: 1902693 · Report as offensive
JohnDK Crowdfunding Project Donor*Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 28 May 00
Posts: 1222
Credit: 451,243,443
RAC: 1,127
Denmark
Message 1902695 - Posted: 24 Nov 2017, 12:11:44 UTC

Don't know why but when I started the PC just now, BOINC did detect both GPUs. Yesterday I rebooted several times with no luck.

Seems to run the cuda8 tasks OK, will try installing cuda9 app...
ID: 1902695 · Report as offensive
TBar
Volunteer tester

Send message
Joined: 22 May 99
Posts: 5204
Credit: 840,779,836
RAC: 2,768
United States
Message 1902708 - Posted: 24 Nov 2017, 15:24:47 UTC - in response to Message 1902695.  

Now that you mention it, I think a couple of us have had the same problem when going back to the repo driver from the nVidia driver in Mint. I think I installed the repo driver 4 times before it found the GPUs. So, that's not uncommon. I've also seen people have problems getting BOINC running again after stopping it in Mint, that was my main reason for going back to Ubuntu on that one machine of mine. Don't be surprised if you also see that problem. I've never had those problems in Ubuntu.
ID: 1902708 · Report as offensive
TBar
Volunteer tester

Send message
Joined: 22 May 99
Posts: 5204
Credit: 840,779,836
RAC: 2,768
United States
Message 1902714 - Posted: 24 Nov 2017, 16:19:26 UTC - in response to Message 1902570.  

Thanks TBar, I was about to shoot you a PM asking about the sigsegv errors with the r3711 app. Never had an issue with either the r3345 or r3306 apps previously. Just installed the made for AMD version. Will continue to monitor.
It's looking good so far, maybe that solved it. Has anyone tried an AVX2 App on those CPUs? It's possible it may work better than the AVX version. It wouldn't be difficult to make the App, just change the SSE41 flags to AVX2. I already have FFTW installed with the AVX2 flag, that's why the AMD App is a little larger than the last SSE41 App, the FFTW library has AVX2 as well as just AVX as in the last App. Of course, someone else would have to test the App.
ID: 1902714 · Report as offensive
JohnDK Crowdfunding Project Donor*Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 28 May 00
Posts: 1222
Credit: 451,243,443
RAC: 1,127
Denmark
Message 1902721 - Posted: 24 Nov 2017, 17:16:00 UTC

I'm up and running the cuda9 app, so far so good :)
ID: 1902721 · Report as offensive
Profile Keith Myers Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 29 Apr 01
Posts: 13161
Credit: 1,160,866,277
RAC: 1,873
United States
Message 1902722 - Posted: 24 Nov 2017, 17:40:45 UTC - in response to Message 1902714.  

Thanks TBar, I was about to shoot you a PM asking about the sigsegv errors with the r3711 app. Never had an issue with either the r3345 or r3306 apps previously. Just installed the made for AMD version. Will continue to monitor.
It's looking good so far, maybe that solved it. Has anyone tried an AVX2 App on those CPUs? It's possible it may work better than the AVX version. It wouldn't be difficult to make the App, just change the SSE41 flags to AVX2. I already have FFTW installed with the AVX2 flag, that's why the AMD App is a little larger than the last SSE41 App, the FFTW library has AVX2 as well as just AVX as in the last App. Of course, someone else would have to test the App.

I can be your guinea pig to try out an AVX2 app. I noticed the larger size of the AMD specific r3711 app compared to the original one. Good to know it was because of the FFTW library change.
Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)
ID: 1902722 · Report as offensive
Profile Jeff Buck Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester

Send message
Joined: 11 Feb 00
Posts: 1441
Credit: 148,764,870
RAC: 0
United States
Message 1902728 - Posted: 24 Nov 2017, 18:08:14 UTC

Here's one that might be worth making an offline run with. It's a case where the Cuda 9 zi3v reported an extra Pulse, which seems like a good thing when it's matched with a Cuda 8 app. However, in this case it was up against a CPU app, so it may not be the better result on this one.

Workunit 2756468548 (blc25_2bit_guppi_57895_47387_HIP91358_0034.24610.818.23.46.191.vlar)
Task 6187707993 (S=5, A=0, P=10, T=3, G=0, BS=26.4327, BG=0) v8.08 (alt) windows_x86_64
Task 6187707994 (S=5, A=0, P=11, T=3, G=0, BS=26.43272, BG=0) x41p_zi3v, Cuda 9.00 special

I'm currently running an offline test on another interesting WU, which will probably take at least a few hours, so I wouldn't have time to run this one until this afternoon, at the earliest.
ID: 1902728 · Report as offensive
Profile Jeff Buck Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester

Send message
Joined: 11 Feb 00
Posts: 1441
Credit: 148,764,870
RAC: 0
United States
Message 1902732 - Posted: 24 Nov 2017, 18:30:37 UTC
Last modified: 24 Nov 2017, 18:40:25 UTC

I also noted something on an Inconclusive due to a "bad" Best Pulse. Yes, I know that's an old issue, which is why I don't think I've ever reported on one, and I've certainly never looked closely at any of them before.

Workunit 2756468564 (blc25_2bit_guppi_57895_47387_HIP91358_0034.24610.818.23.46.193.vlar)
Task 6187708167 (S=0, A=0, P=7, T=1, G=0, BS=23.86436, BG=0) x41p_zi3v, Cuda 9.00 special
Task 6187708168 (S=0, A=0, P=7, T=1, G=0, BS=23.86423, BG=0) v8.08 (alt) windows_x86_64

Each app identified a different one of the 7 reported Pulses as Best. Zeroing in on those 2 Pulses, the Special App chose the first one of these two:
Pulse: peak=4.518389, time=45.9, period=10.26, d_freq=2082869054.8, score=1.042, chirp=-71.733, fft_len=2k
Pulse: peak=8.279277, time=45.84, period=18.76, d_freq=2082875244.64, score=1.041, chirp=96.855, fft_len=512

...whereas the CPU app chose the second one of these two:
Pulse: peak=4.518386, time=45.9, period=10.26, d_freq=2082869054.8, score=1.041, chirp=-71.733, fft_len=2k
Pulse: peak=8.279298, time=45.84, period=18.76, d_freq=2082875244.64, score=1.041, chirp=96.855, fft_len=512

To me, it looks like a simple precision/rounding issue in the Pulse's "score" value. The Special App picked the one with the barely perceptible higher score, while the CPU app picked the last of two with apparently identical scores.

The thing is, from a "score" standpoint, I would expect either of these two Pulses to be acceptable to the validator as the Best Pulse. So I'm guessing that, although "score" is used to choose the Best Pulse, perhaps the validator is using the "peak" value to decide whether or not they match. Can anybody confirm that?

EDIT: And then, too, is this WU representative of other WUs where the so-called "bad" Best Pulse issue has surfaced with the Special App? Or was it just pure luck that I happened to notice one with nearly identical score values? (As I said, I hadn't really been paying much attention to these.)
ID: 1902732 · Report as offensive
Profile Jeff Buck Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester

Send message
Joined: 11 Feb 00
Posts: 1441
Credit: 148,764,870
RAC: 0
United States
Message 1902780 - Posted: 24 Nov 2017, 22:18:18 UTC
Last modified: 24 Nov 2017, 22:23:08 UTC

A couple months ago I reported (in Message 1889868) on an Inconclusive where one host reported 30 signals (a -9 overflow) and the other reported 29 signals (not an overflow) . This morning I spotted another one, and this ties into the observation that the Cuda9 zi3v appears to be correctly finding some Pulses that other apps have missed. (Yes, I know the other host in this example is an Intel GPU, but I'm not sure that's relevant to this situation, unless it actually missed more than the single Pulse.)

Workunit 2756518670 (03mr07aa.18236.13978.15.42.121)
Task 6187812028 (S=25, A=2, P=3, T=0, G=0, BS=27.79362, BG=4.143701) x41p_zi3v, Cuda 9.00 special
Task 6187812029 (S=25, A=2, P=2, T=0, G=0, BS=27.79354, BG=4.143687) v8.00 (opencl_intel_gpu_sah) x86_64-apple-darwin

I just finished confirming the Cuda9 zi3v results with an offline Windows CPU run. Assuming that the tiebreaker for this WU does the same, this means that the zi3v will likely be designated the canonical result, as it should be. But.......that one extra Pulse changes the result from non-overflow to overflow, simply because it now has 30 signals instead of 29. Does that make it worthless, inasmuch as it's become an overflow because it's more accurate? I don't see how, unless a higher cut-off would somehow also result in more than 30 reported signals from the zi3v and CPU apps. It doesn't seem likely, but would be difficult to know for sure. I think that the only possibility would be to run this WU through some sort of modified app with a cut-off higher than 30 built in. That's something that a developer could perhaps take a crack at.

EDIT: Added a link to my earlier post.
ID: 1902780 · Report as offensive
TBar
Volunteer tester

Send message
Joined: 22 May 99
Posts: 5204
Credit: 840,779,836
RAC: 2,768
United States
Message 1902795 - Posted: 25 Nov 2017, 1:43:40 UTC

Hey, the New Repository version of 384.90 seems to have solved the problem my machines were having with Kernel 4.10 and the nVidia driver 384. With the repository driver it now works in Ubuntu 16.04 & 17.04, nice.
ID: 1902795 · Report as offensive
Profile Jeff Buck Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester

Send message
Joined: 11 Feb 00
Posts: 1441
Credit: 148,764,870
RAC: 0
United States
Message 1902796 - Posted: 25 Nov 2017, 1:57:07 UTC - in response to Message 1902728.  

Here's one that might be worth making an offline run with. It's a case where the Cuda 9 zi3v reported an extra Pulse, which seems like a good thing when it's matched with a Cuda 8 app. However, in this case it was up against a CPU app, so it may not be the better result on this one.

Workunit 2756468548 (blc25_2bit_guppi_57895_47387_HIP91358_0034.24610.818.23.46.191.vlar)
Task 6187707993 (S=5, A=0, P=10, T=3, G=0, BS=26.4327, BG=0) v8.08 (alt) windows_x86_64
Task 6187707994 (S=5, A=0, P=11, T=3, G=0, BS=26.43272, BG=0) x41p_zi3v, Cuda 9.00 special

I'm currently running an offline test on another interesting WU, which will probably take at least a few hours, so I wouldn't have time to run this one until this afternoon, at the earliest.
Okay, i finally got around to making an offline run for this one, using the stock Windows CPU app (setiathome_8.00_windows_intelx86). It matches the Cuda9 zi3v result. The Pulse that v8.08 (alt) didn't report was:

Pulse: peak=6.338756, time=45.9, period=18.07, d_freq=2082851430.83, score=1, chirp=27.732, fft_len=2k

Could the fact that the score value was exactly "1" (at least as reported by zi3v) offer an explanation? The result file from the offline run just has a value of "0" for the score.
ID: 1902796 · Report as offensive
Profile Raistmer
Volunteer developer
Volunteer tester
Avatar

Send message
Joined: 16 Jun 01
Posts: 6325
Credit: 106,370,077
RAC: 121
Russia
Message 1902914 - Posted: 25 Nov 2017, 21:38:08 UTC - in response to Message 1902732.  


To me, it looks like a simple precision/rounding issue in the Pulse's "score" value. The Special App picked the one with the barely perceptible higher score, while the CPU app picked the last of two with apparently identical scores.

Seems so. At leastPEtri's app choice is logical with given score it computed.

[quote
The thing is, from a "score" standpoint, I would expect either of these two Pulses to be acceptable to the validator as the Best Pulse. So I'm guessing that, although "score" is used to choose the Best Pulse, perhaps the validator is using the "peak" value to decide whether or not they match. Can anybody confirm that?
[/quote]
Validator looks for many properties of reported pulse, peak and its position including.


EDIT: And then, too, is this WU representative of other WUs where the so-called "bad" Best Pulse issue has surfaced with the Special App? Or was it just pure luck that I happened to notice one with nearly identical score values? (As I said, I hadn't really been paying much attention to these.)

[/quote]
No, this caseisn't representative. Best often one of reported but usually this issue shows itself w/o such good correlation with scores.
SETI apps news
We're not gonna fight them. We're gonna transcend them.
ID: 1902914 · Report as offensive
Profile Raistmer
Volunteer developer
Volunteer tester
Avatar

Send message
Joined: 16 Jun 01
Posts: 6325
Credit: 106,370,077
RAC: 121
Russia
Message 1902915 - Posted: 25 Nov 2017, 21:38:10 UTC - in response to Message 1902732.  


To me, it looks like a simple precision/rounding issue in the Pulse's "score" value. The Special App picked the one with the barely perceptible higher score, while the CPU app picked the last of two with apparently identical scores.

Seems so. At leastPEtri's app choice is logical with given score it computed.

[quote
The thing is, from a "score" standpoint, I would expect either of these two Pulses to be acceptable to the validator as the Best Pulse. So I'm guessing that, although "score" is used to choose the Best Pulse, perhaps the validator is using the "peak" value to decide whether or not they match. Can anybody confirm that?
[/quote]
Validator looks for many properties of reported pulse, peak and its position including.


EDIT: And then, too, is this WU representative of other WUs where the so-called "bad" Best Pulse issue has surfaced with the Special App? Or was it just pure luck that I happened to notice one with nearly identical score values? (As I said, I hadn't really been paying much attention to these.)

[/quote]
No, this caseisn't representative. Best often one of reported but usually this issue shows itself w/o such good correlation with scores.
SETI apps news
We're not gonna fight them. We're gonna transcend them.
ID: 1902915 · Report as offensive
Profile Raistmer
Volunteer developer
Volunteer tester
Avatar

Send message
Joined: 16 Jun 01
Posts: 6325
Credit: 106,370,077
RAC: 121
Russia
Message 1902917 - Posted: 25 Nov 2017, 21:44:44 UTC - in response to Message 1902796.  

Here's one that might be worth making an offline run with. It's a case where the Cuda 9 zi3v reported an extra Pulse, which seems like a good thing when it's matched with a Cuda 8 app. However, in this case it was up against a CPU app, so it may not be the better result on this one.

Workunit 2756468548 (blc25_2bit_guppi_57895_47387_HIP91358_0034.24610.818.23.46.191.vlar)
Task 6187707993 (S=5, A=0, P=10, T=3, G=0, BS=26.4327, BG=0) v8.08 (alt) windows_x86_64
Task 6187707994 (S=5, A=0, P=11, T=3, G=0, BS=26.43272, BG=0) x41p_zi3v, Cuda 9.00 special

I'm currently running an offline test on another interesting WU, which will probably take at least a few hours, so I wouldn't have time to run this one until this afternoon, at the earliest.
Okay, i finally got around to making an offline run for this one, using the stock Windows CPU app (setiathome_8.00_windows_intelx86). It matches the Cuda9 zi3v result. The Pulse that v8.08 (alt) didn't report was:

Pulse: peak=6.338756, time=45.9, period=18.07, d_freq=2082851430.83, score=1, chirp=27.732, fft_len=2k

Could the fact that the score value was exactly "1" (at least as reported by zi3v) offer an explanation? The result file from the offline run just has a value of "0" for the score.


Could you give download link to this task (it removed from SETI server already).

BTW, OpenCL apps have additional featureto print threshold valuesforpulses if -v 2 switch provided. Useful for offline tasks .
SETI apps news
We're not gonna fight them. We're gonna transcend them.
ID: 1902917 · Report as offensive
Profile Jeff Buck Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester

Send message
Joined: 11 Feb 00
Posts: 1441
Credit: 148,764,870
RAC: 0
United States
Message 1902922 - Posted: 25 Nov 2017, 22:47:55 UTC - in response to Message 1902917.  

Could you give download link to this task (it removed from SETI server already).
Here you go: https://www.dropbox.com/s/3i4763b26c463xc/WU%202756468548.7z?dl=0
File contains WU and output from my Windows CPU offline run.

And thanks for your reply about the Best Pulse example. I was hoping there might be a clue in that one, but apparently not. :^)
ID: 1902922 · Report as offensive
Profile Jeff Buck Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester

Send message
Joined: 11 Feb 00
Posts: 1441
Credit: 148,764,870
RAC: 0
United States
Message 1902950 - Posted: 26 Nov 2017, 3:37:23 UTC

I don't believe I've seen this type of Inconclusive previously. The v8.08 (alt) Windows CPU app reported a Triplet that the Special App doesn't seem to have recognized at all. Even the Best Triplet reported by the Special App shows "peak=0".

Workunit 2757131579 (blc25_2bit_guppi_57895_51266_HIP2_0039.24131.818.24.47.42.vlar)
Task 6189093728 (S=6, A=1, P=6, T=0, G=0, BS=26.20532, BG=0) x41p_zi3v, Cuda 9.00 special
Task 6189093729 (S=6, A=1, P=6, T=1, G=0, BS=26.20529, BG=0) v8.08 (alt) windows_x86_64

The missing signal is:
Triplet: peak=10.93743, time=34.56, period=20.22, d_freq=2201264522.54, chirp=-14.715, fft_len=512

It looks like the tiebreaker will also run with v8.08 (alt), so it would be helpful if someone can run this WU with one or more other apps. I can probably run it with the stock Windows CPU app tomorrow but, theoretically, that should agree with the alternate Windows CPU app.
ID: 1902950 · Report as offensive
Previous · 1 . . . 71 · 72 · 73 · 74 · 75 · 76 · 77 . . . 83 · Next

Message boards : Number crunching : Linux CUDA 'Special' App finally available, featuring Low CPU use


 
©2024 University of California
 
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.