SETI@home v8.12 Windows GPU applications support thread

Message boards : Number crunching : SETI@home v8.12 Windows GPU applications support thread
Message board moderation

To post messages, you must log in.

Previous · 1 · 2 · 3 · 4 · 5 · 6 · 7 . . . 17 · Next

AuthorMessage
Profile Raistmer
Volunteer developer
Volunteer tester
Avatar

Send message
Joined: 16 Jun 01
Posts: 6016
Credit: 85,727,937
RAC: 35,691
Russia
Message 1800134 - Posted: 2 Jul 2016, 16:05:45 UTC - in response to Message 1800132.  
Last modified: 2 Jul 2016, 16:06:14 UTC


Both GUPPI and non-guppi have increased time to complete. I run 3 at a time and total increase is about 4 minutes for both groups.

And what about CPU time consumption?
And do you run sleep enabled or disabled?
SETI apps news
We're not gonna fight them. We're gonna transcend them.
ID: 1800134 · Report as offensive
Profile Zalster Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 27 May 99
Posts: 4763
Credit: 310,426,109
RAC: 340,699
United States
Message 1800136 - Posted: 2 Jul 2016, 16:08:11 UTC - in response to Message 1800134.  

I run without the -use_sleep as these are pure crunchers and large CPU chip

Sorry I have not checked CPU time consumption, I would need to go back and look at that but it wasn't something that I was concerned with.
ID: 1800136 · Report as offensive
Profile Raistmer
Volunteer developer
Volunteer tester
Avatar

Send message
Joined: 16 Jun 01
Posts: 6016
Credit: 85,727,937
RAC: 35,691
Russia
Message 1800139 - Posted: 2 Jul 2016, 16:26:34 UTC - in response to Message 1800136.  

I run without the -use_sleep as these are pure crunchers and large CPU chip

Sorry I have not checked CPU time consumption, I would need to go back and look at that but it wasn't something that I was concerned with.

Well, it's the question of balance between overall performance, CPU consumption, usability for both high-, mid- and low-end devices. Since 8.12 no single GPU kernel was changed. Only way they enqueued.
There are different approaches as you could see. Some dont' care about CPU consumption, some care only about CPU consumption no matter how good overall throughput is...
SETI apps news
We're not gonna fight them. We're gonna transcend them.
ID: 1800139 · Report as offensive
Richard Haselgrove Project Donor
Volunteer tester

Send message
Joined: 4 Jul 99
Posts: 12271
Credit: 125,965,813
RAC: 40,113
United Kingdom
Message 1800140 - Posted: 2 Jul 2016, 16:35:35 UTC - in response to Message 1800139.  

no matter how good overall throughput is...

Remember to draw a distinction between overall throughput and SETI throughput...
ID: 1800140 · Report as offensive
Stephen "Heretic" Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 20 Sep 12
Posts: 3769
Credit: 88,532,074
RAC: 159,467
Australia
Message 1800234 - Posted: 3 Jul 2016, 1:10:14 UTC - in response to Message 1800139.  

I run without the -use_sleep as these are pure crunchers and large CPU chip

Sorry I have not checked CPU time consumption, I would need to go back and look at that but it wasn't something that I was concerned with.

Well, it's the question of balance between overall performance, CPU consumption, usability for both high-, mid- and low-end devices. Since 8.12 no single GPU kernel was changed. Only way they enqueued.
There are different approaches as you could see. Some dont' care about CPU consumption, some care only about CPU consumption no matter how good overall throughput is...


. . On that thought, does -use_sleep_ex 0 mean turn off use sleep or use less than a value of 1? I am considering trying the latest version on the 970's to see how they behave with it.
ID: 1800234 · Report as offensive
Profile Zalster Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 27 May 99
Posts: 4763
Credit: 310,426,109
RAC: 340,699
United States
Message 1800568 - Posted: 4 Jul 2016, 4:18:30 UTC

Raistmer,

Just going to throw this out there.

I know you are concentrating on getting the newer SoG to work with the lower end GPUs, but I find r3430 faster on the higher end GPUs.

The only issue is that sanity check that is causing errors with that version.

Any chance of a patch for r3430 so we can eliminate those errors?
ID: 1800568 · Report as offensive
rob smith Special Project $250 donor
Volunteer tester

Send message
Joined: 7 Mar 03
Posts: 16578
Credit: 345,538,647
RAC: 230,825
United Kingdom
Message 1800629 - Posted: 4 Jul 2016, 10:55:52 UTC

From a very simplistic and unscientific observation of three of my crunchers - The error rate appears to increase as the period of time from last re-boot increases.

Normally I would associate this with a memory leak, or memory not being cleared/released properly, but I'm not sure if this applies to GPU systems.
Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?
ID: 1800629 · Report as offensive
Profile Stubbles
Volunteer tester
Avatar

Send message
Joined: 29 Nov 99
Posts: 358
Credit: 5,909,255
RAC: 0
Canada
Message 1800780 - Posted: 4 Jul 2016, 21:30:54 UTC

On one of my two almost-identical rigs (HP Z400 with Xeon W3550 and GTX 750 Ti), I am running Lunatics v0.45 beta3 with SoG (MB8_win_x86_SSE3_OpenCL_NV_SoG_r3472.exe) and crunching 1 WU/GPU.

With BoinStats, I noticed a few tasks are not fully utilizing the dedicated CPU core to support the GPU. Surprisingly, those tasks are finishing much more quickly (in almost half the time).

"CPU %" = ~15% (as per BoincStats)
SETI@home	8.12 setiathome_v8 (opencl_nvidia_SoG)	21au10ah.4470.3343.9.36.177_0	00:08:47 (00:01:17)	2016-Jul-04 8:45:38 AM	2016-Jul-04 9:44:14 AM	14.61
SETI@home	8.12 setiathome_v8 (opencl_nvidia_SoG)	21au10ah.4470.3343.9.36.186_1	00:08:46 (00:01:16)	2016-Jul-04 8:53:41 AM	2016-Jul-04 9:44:14 AM	14.45
SETI@home	8.12 setiathome_v8 (opencl_nvidia_SoG)	21au10ah.4470.3343.9.36.183_0	00:08:47 (00:01:18)	2016-Jul-04 9:02:46 AM	2016-Jul-04 9:44:14 AM	14.80
SETI@home	8.12 setiathome_v8 (opencl_nvidia_SoG)	21au10ah.4470.3343.9.36.220_0	00:08:47 (00:01:18)	2016-Jul-04 9:29:35 AM	2016-Jul-04 9:44:14 AM	14.80

usual CPU % activity
SETI@home	8.12 setiathome_v8 (opencl_nvidia_SoG)	21au10ah.507.36467.10.37.65_0	00:16:44 (00:16:36)	2016-Jul-04 10:09:19 AM	2016-Jul-04 10:24:42 AM	99.20
SETI@home	8.12 setiathome_v8 (opencl_nvidia_SoG)	21au10ah.507.36467.10.37.25_0	00:16:36 (00:16:31)	2016-Jul-04 10:24:20 AM	2016-Jul-04 10:24:42 AM	99.50
SETI@home	8.12 setiathome_v8 (opencl_nvidia_SoG)	21ap10af.4505.9065.12.39.164_0	00:16:33 (00:16:39)	2016-Jul-04 9:47:14 AM	2016-Jul-04 9:59:40 AM	100.00
Is there a reason for this?

Fyi#1, I've been moving non-vlars from CPU to GPU but I haven't kept track of the ones that I did move...but I could figure that out for future one by comparing them to the: "In progress" list before BoincStats reports them.

Fyi#2, during the last week, I inserting supplied parameters into app_config.XML and mb_cmdline_win_x86_SSE3_OpenCL_NV.txt using Brent Norman's suggestion.

If there's any testing I could do, let me know the parameters.
Rob :-}
ID: 1800780 · Report as offensive
Profile Zalster Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 27 May 99
Posts: 4763
Credit: 310,426,109
RAC: 340,699
United States
Message 1800791 - Posted: 4 Jul 2016, 21:58:14 UTC - in response to Message 1800780.  
Last modified: 4 Jul 2016, 21:59:30 UTC

Rob,

I can't speak for Raistmer, but I can speak from my understanding on this.

When the new work units from Green Bank were released, the work units were much longer (taking the appearance of VLAR) and as such were relegated to the CPUs.

Newer cuda apps (v8) were needed to handle this new data. Eventually we got those but found the times were extremely long compared to the previous work units.(but less than CPU times) Also, the amount of credit for these work units decreased.

This movement to take work units from the GPU back to CPU is a response to that change.(credit wise)

SoG and Sah were created to find a faster way to analyze the new data. And despite what you hear, it does just that.

In short what I'm saying is, if you want to help with results, you would need to leave those work units assigned to the GPU to remain there and run them on these apps. Otherwise, you won't be helping by reassigning work units.

Sorry if I sound like I'm harping but since very few have volunteered their machines to help him figure out the issues, it's not surprising that his apps work better for those who lent their machine than those that sat on the sidelines.

Help is always welcome.
ID: 1800791 · Report as offensive
Profile Raistmer
Volunteer developer
Volunteer tester
Avatar

Send message
Joined: 16 Jun 01
Posts: 6016
Credit: 85,727,937
RAC: 35,691
Russia
Message 1800795 - Posted: 4 Jul 2016, 22:04:48 UTC - in response to Message 1800568.  


Any chance of a patch for r3430 so we can eliminate those errors?

8.15 will be released soon. I expect back to full speed for high-end devices. Will see.
SETI apps news
We're not gonna fight them. We're gonna transcend them.
ID: 1800795 · Report as offensive
Profile Raistmer
Volunteer developer
Volunteer tester
Avatar

Send message
Joined: 16 Jun 01
Posts: 6016
Credit: 85,727,937
RAC: 35,691
Russia
Message 1800796 - Posted: 4 Jul 2016, 22:07:57 UTC - in response to Message 1800629.  
Last modified: 4 Jul 2016, 22:08:21 UTC

From a very simplistic and unscientific observation of three of my crunchers - The error rate appears to increase as the period of time from last re-boot increases.

Normally I would associate this with a memory leak, or memory not being cleared/released properly, but I'm not sure if this applies to GPU systems.

About what error rate you talk? What app's flavor? ATi, NV, iGPU ?
Flase positive on overflow with many high-power autocorr has nothing to do with any leak or app bug. It's safeguard measure against bad-behaving GPU devices. With new data its threshold exceeded for genuine operation too hence false positive. It's data-driven, not machine state.
SETI apps news
We're not gonna fight them. We're gonna transcend them.
ID: 1800796 · Report as offensive
Profile Raistmer
Volunteer developer
Volunteer tester
Avatar

Send message
Joined: 16 Jun 01
Posts: 6016
Credit: 85,727,937
RAC: 35,691
Russia
Message 1800797 - Posted: 4 Jul 2016, 22:09:17 UTC - in response to Message 1800780.  

On one of my two almost-identical rigs (HP Z400 with Xeon W3550 and GTX 750 Ti), I am running Lunatics v0.45 beta3 with SoG (MB8_win_x86_SSE3_OpenCL_NV_SoG_r3472.exe) and crunching 1 WU/GPU.

....

What issue here?
SETI apps news
We're not gonna fight them. We're gonna transcend them.
ID: 1800797 · Report as offensive
J. Mileski
Volunteer tester
Avatar

Send message
Joined: 9 Jun 02
Posts: 563
Credit: 130,088,397
RAC: 8,423
United States
Message 1800798 - Posted: 4 Jul 2016, 22:12:09 UTC

I'm willing to move a machine or 2 to beta. What kind of hardware are you looking for? I've got GTX550ti, GTX750ti GT240, 9500GT, and 9400GT
ID: 1800798 · Report as offensive
Profile Raistmer
Volunteer developer
Volunteer tester
Avatar

Send message
Joined: 16 Jun 01
Posts: 6016
Credit: 85,727,937
RAC: 35,691
Russia
Message 1800802 - Posted: 4 Jul 2016, 22:31:43 UTC - in response to Message 1800798.  

I'm willing to move a machine or 2 to beta. What kind of hardware are you looking for? I've got GTX550ti, GTX750ti GT240, 9500GT, and 9400GT

Currently beta tests new data format so any will go.
For testing OpenCL NV apps smth with CC 2.0+ (FERMI and up) will go.
SETI apps news
We're not gonna fight them. We're gonna transcend them.
ID: 1800802 · Report as offensive
Profile Stubbles
Volunteer tester
Avatar

Send message
Joined: 29 Nov 99
Posts: 358
Credit: 5,909,255
RAC: 0
Canada
Message 1800807 - Posted: 4 Jul 2016, 22:57:15 UTC - in response to Message 1800797.  

On one of my two almost-identical rigs (HP Z400 with Xeon W3550 and GTX 750 Ti), I am running Lunatics v0.45 beta3 with SoG (MB8_win_x86_SSE3_OpenCL_NV_SoG_r3472.exe) and crunching 1 WU/GPU.
....
What issue here?
There isn't an issue; just an observation.
I asked:
Is there a reason for this?
... 1) diff in CPU support times, and
2) why does ~15% CPU support on SoG (instead of ~100%) make the task run in half the time.
ID: 1800807 · Report as offensive
J. Mileski
Volunteer tester
Avatar

Send message
Joined: 9 Jun 02
Posts: 563
Credit: 130,088,397
RAC: 8,423
United States
Message 1800813 - Posted: 4 Jul 2016, 23:16:17 UTC - in response to Message 1800802.  

I'm willing to move a machine or 2 to beta. What kind of hardware are you looking for? I've got GTX550ti, GTX750ti GT240, 9500GT, and 9400GT

Currently beta tests new data format so any will go.
For testing OpenCL NV apps smth with CC 2.0+ (FERMI and up) will go.

For the Open CL NV apps, are they being tested here or over at beta?
ID: 1800813 · Report as offensive
Profile Stubbles
Volunteer tester
Avatar

Send message
Joined: 29 Nov 99
Posts: 358
Credit: 5,909,255
RAC: 0
Canada
Message 1800814 - Posted: 4 Jul 2016, 23:21:23 UTC - in response to Message 1800791.  

In short what I'm saying is, if you want to help with results, you would need to leave those work units assigned to the GPU to remain there and run them on these apps. Otherwise, you won't be helping by reassigning work units.
Sorry if I sound like I'm harping but since very few have volunteered their machines to help him figure out the issues, it's not surprising that his apps work better for those who lent their machine than those that sat on the sidelines.
Help is always welcome.


Thanks for the direct feedback.
In the 2 months that I have spent learning from the S@H forums, I have learned a lot! I only just started reassigning tasks during the last few days while waiting for Beta4.
Also, it's only today from my own digging that I realized how few are actually using Lunatics optimized apps. I assumed most of the top 5,000 computers were!

If I understand correctly (correct me if I'm wrong), there's 2 types of "volunteer testers":
1. those who simply crunch tasks for the "S@H Beta Test" project with the stock apps in the same way that anyone does for the regular S@H project; and
2. those who test the latest Lunatic apps on the regular: S@H project.

Also, is there a newer app floating somewhere other than SoG_r3472.exe that I can run the guppis on?
If so, how many WU on a GTX 750 Ti? 1 or 2 in parallel
I'll need to have command line parameters supplied to me if there are any specific ones to test...as I haven't started to learn anything yet about the diff options...but I feel I'm somewhat ready to learn & willing!

Happy 4th Zalster
ID: 1800814 · Report as offensive
Profile Zalster Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 27 May 99
Posts: 4763
Credit: 310,426,109
RAC: 340,699
United States
Message 1800821 - Posted: 4 Jul 2016, 23:42:25 UTC - in response to Message 1800814.  

Hey Rob,

Actually, they would have preferred all testing to have been done on Beta.

The fact that some of us are running the test apps on Main is being allowed by Raistmer to see how they work with the data.

On the beta site, while some people do use commandlines and app_config, it is preferable that applications be run as is, without any modifications so they can review the results and see where they need to improve the apps.

Some people, and I include myself here, do add commmadlines and app_config but only after running a large amount of work units under stock conditions to make sure we have a baseline.

I use 100 validated results, not inconclusive, not pending. So that might take a while.

I add commandlines to try and spend up the times to complete as well as find the best ratio for my machines. Now, that doesn't always translate into a direct swap from beta to main as I pointed out above and asked about.

I'm sure Rasitmer can give you more guidelines in what he wants in testing his apps.

Happy 1st of July to you as well.

Zalster
ID: 1800821 · Report as offensive
rob smith Special Project $250 donor
Volunteer tester

Send message
Joined: 7 Mar 03
Posts: 16578
Credit: 345,538,647
RAC: 230,825
United Kingdom
Message 1800874 - Posted: 5 Jul 2016, 5:13:25 UTC - in response to Message 1800796.  

All three of my current PCs are NV based. Each of the PCs has a pair of GPU, one with GTX780, one with GTX960 and one with GTX980s
It would appear that in the first few hours of running after a reboot I get no errors, then one in the next few hours, and progressively get more per hour (actually fewer hours between errors).
Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?
ID: 1800874 · Report as offensive
Profile Raistmer
Volunteer developer
Volunteer tester
Avatar

Send message
Joined: 16 Jun 01
Posts: 6016
Credit: 85,727,937
RAC: 35,691
Russia
Message 1800882 - Posted: 5 Jul 2016, 5:41:34 UTC - in response to Message 1800807.  

On one of my two almost-identical rigs (HP Z400 with Xeon W3550 and GTX 750 Ti), I am running Lunatics v0.45 beta3 with SoG (MB8_win_x86_SSE3_OpenCL_NV_SoG_r3472.exe) and crunching 1 WU/GPU.
....
What issue here?
There isn't an issue; just an observation.
I asked:
Is there a reason for this?
... 1) diff in CPU support times, and
2) why does ~15% CPU support on SoG (instead of ~100%) make the task run in half the time.

Did you check AR values for tasks in comparison? Different AR means different set of searches used so different performance and behavior.
SETI apps news
We're not gonna fight them. We're gonna transcend them.
ID: 1800882 · Report as offensive
Previous · 1 · 2 · 3 · 4 · 5 · 6 · 7 . . . 17 · Next

Message boards : Number crunching : SETI@home v8.12 Windows GPU applications support thread


 
©2018 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.