SETI@home v8.22 Windows GPU applications support thread

Message boards : Number crunching : SETI@home v8.22 Windows GPU applications support thread
Message board moderation

To post messages, you must log in.

Previous · 1 . . . 13 · 14 · 15 · 16 · 17 · 18 · 19 . . . 22 · Next

AuthorMessage
Profile Mike Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 17 Feb 01
Posts: 31825
Credit: 74,547,470
RAC: 24,676
Germany
Message 1951415 - Posted: 22 Aug 2018, 22:01:46 UTC - in response to Message 1951387.  

Has anyone been playing with this setting -pref_wg_num_per_cu ?


Of course as the alpha tester of this app i did.
And what is the question ?
With each crime and every kindness we birth our future.
ID: 1951415 · Report as offensive     Reply Quote
Marco Vandebergh

Send message
Joined: 27 Aug 10
Posts: 37
Credit: 11,156,309
RAC: 7,920
Netherlands
Message 1951621 - Posted: 23 Aug 2018, 19:20:02 UTC - in response to Message 1951415.  

What are your settings with the -pref_wg_num_per_cu?

Im getting quite an improvement on my 1080ti increasing this setting, especially when running for example 3 of 4 units at once in combination with cpu_lock.

I did not find much information about it, so i thought id ask.
ID: 1951621 · Report as offensive     Reply Quote
Profile Mike Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 17 Feb 01
Posts: 31825
Credit: 74,547,470
RAC: 24,676
Germany
Message 1951654 - Posted: 23 Aug 2018, 20:52:42 UTC

I don`t use this app arg because my GPU don`t get an improvement from it.
Sadly i had no chance to test it on a 1080.
It certainly would be interesting to see how it performs in high production mode.
With each crime and every kindness we birth our future.
ID: 1951654 · Report as offensive     Reply Quote
Marco Vandebergh

Send message
Joined: 27 Aug 10
Posts: 37
Credit: 11,156,309
RAC: 7,920
Netherlands
Message 1951664 - Posted: 23 Aug 2018, 21:13:41 UTC - in response to Message 1951654.  

Ok, ill leave this running for a while, to see how it perfoms on the longer run.

In GPU-Z the GPU load is now around 96 to 99 % and memory controller load about 86 / 91 ish.
It was really low before when running with standard parameters, even when used 3 WU's at a time.


I do have to mention the GPU gets hotter, so becareful. My system is all watercooled i have no issues. max GPU temp is 46 degree C.


my app_config:

<app_config>
<app_version>
<app_name>setiathome_v8</app_name>
<plan_class>opencl_nvidia_SoG</plan_class>
<avg_ncpus>1</avg_ncpus>
<ngpus>0.33</ngpus>
<cmdline>-sbs 1024 -total_GPU_instances_num 5 -pref_wg_num_per_cu 20 -cpu_lock -period_iterations_num 1 -hp -high_prec_timer -high_perf -tt 1500</cmdline>
</app_version>
<app_version>
<app_name>astropulse_v7</app_name>
<plan_class>opencl_nvidia_100</plan_class>
<avg_ncpus>1</avg_ncpus>
<ngpus>1</ngpus>
<cmdline>-unroll 28 -oclFFT_plan 256 16 256 -ffa_block 12288 -ffa_block_fetch 6144 -tune 1 64 4 1 -tune 2 64 4 1 -hp -instances_per_device 1</cmdline>
</app_version>
</app_config>
ID: 1951664 · Report as offensive     Reply Quote
Profile ab1vl-pi
Avatar

Send message
Joined: 11 Jun 18
Posts: 4
Credit: 68,014
RAC: 0
United States
Message 1952063 - Posted: 25 Aug 2018, 21:32:38 UTC

Looking for Fermi or later GPU Nvidia hardware.

As I understand it only ASTROPULSE cares about buggy post 340.x drivers for pre-fermi hardware so I might be open to that too.

I have a handful of ASUS mobos with Athlon 64 2x and 4x cores. The 2X with Gt-220 GPU runs blc04_2bit_guppi WUs (showing as v8.8.01 cuda60)on linux and win7, same box.

If you have a used GPU you'd like to sell, PCI or PCIe, that is faster than the 1.2 cc the GT-220 offers, I'm interested in talking.

If you can recommend a low cost NEW card that offers decent performance, I'd like to hear your recommendations. I don't game, I just want a few decent co-processors to stuff into these systems I rescued from the transfer center.

Please use PM.
tnx, chuck
ID: 1952063 · Report as offensive     Reply Quote
Grant (SSSF)
Volunteer tester

Send message
Joined: 19 Aug 99
Posts: 11559
Credit: 170,142,294
RAC: 105,301
Australia
Message 1952064 - Posted: 25 Aug 2018, 21:40:46 UTC - in response to Message 1952063.  

If you can recommend a low cost NEW card that offers decent performance, I'd like to hear your recommendations. I don't game, I just want a few decent co-processors to stuff into these systems I rescued from the transfer center.

For cost, performance and power use it's hard to go past the GTX 750Ti.
Grant
Darwin NT
ID: 1952064 · Report as offensive     Reply Quote
Stephen "Heretic" Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 20 Sep 12
Posts: 4621
Credit: 143,154,646
RAC: 245,398
Australia
Message 1952076 - Posted: 25 Aug 2018, 23:40:57 UTC - in response to Message 1951664.  
Last modified: 25 Aug 2018, 23:41:40 UTC

Ok, ill leave this running for a while, to see how it perfoms on the longer run.

In GPU-Z the GPU load is now around 96 to 99 % and memory controller load about 86 / 91 ish.
It was really low before when running with standard parameters, even when used 3 WU's at a time.


I do have to mention the GPU gets hotter, so becareful. My system is all watercooled i have no issues. max GPU temp is 46 degree C.


my app_config:

<app_config>
<avg_ncpus>1</avg_ncpus>
<ngpus>0.33</ngpus>
<cmdline>-sbs 1024 -total_GPU_instances_num 5 -pref_wg_num_per_cu 20 -cpu_lock -period_iterations_num 1 -hp -high_prec_timer -high_perf -tt 1500</cmdline>

</app_config>

. . So you are still running 3 at a time? With the change in settings how does it utilise the GPU running one or two at a time?

Stephen

??
ID: 1952076 · Report as offensive     Reply Quote
Profile Zalster Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 27 May 99
Posts: 5177
Credit: 430,629,465
RAC: 582,539
United States
Message 1952092 - Posted: 26 Aug 2018, 0:20:43 UTC - in response to Message 1952076.  

If he's on windows running SoG on 1080Ti then running 2 at a time is slower than running 3 at a time (average time)
ID: 1952092 · Report as offensive     Reply Quote
Stephen "Heretic" Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 20 Sep 12
Posts: 4621
Credit: 143,154,646
RAC: 245,398
Australia
Message 1952136 - Posted: 26 Aug 2018, 4:13:03 UTC - in response to Message 1952092.  

If he's on windows running SoG on 1080Ti then running 2 at a time is slower than running 3 at a time (average time)


. . But he is running a parameter I haven't heard anything about before -

-pref_wg_num_per_cu 20

and he says it increases the GPU utilisation. I was wondering if it might increase it to the max even when running 2 or maybe even one. That would be useful.

Stephen

? ?
ID: 1952136 · Report as offensive     Reply Quote
Profile ab1vl-pi
Avatar

Send message
Joined: 11 Jun 18
Posts: 4
Credit: 68,014
RAC: 0
United States
Message 1952243 - Posted: 26 Aug 2018, 20:38:41 UTC - in response to Message 1952064.  

Tnx Grant, Amazon has one by NXDA (China) labelled as 2BG GTX 750Ti, for $70.90, about 1/5 of the "pure" NVIDIA price.

Any intelligence on NXDA? This would be fulfilled from China ARO 30-45 days.

Another concern is trade wars, Amazon boilerplate leaves us open to an new duties.

Other than that I'd give it a shot. Unless it was DOA it would probably be a big improvement.
ID: 1952243 · Report as offensive     Reply Quote
Profile Zalster Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 27 May 99
Posts: 5177
Credit: 430,629,465
RAC: 582,539
United States
Message 1952252 - Posted: 26 Aug 2018, 21:03:42 UTC - in response to Message 1952136.  



-pref_wg_num_per_cu 20


It's in the read me file of the lunatics installer
ID: 1952252 · Report as offensive     Reply Quote
Stephen "Heretic" Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 20 Sep 12
Posts: 4621
Credit: 143,154,646
RAC: 245,398
Australia
Message 1952264 - Posted: 26 Aug 2018, 22:33:29 UTC - in response to Message 1952252.  



-pref_wg_num_per_cu 20


It's in the read me file of the lunatics installer


. . Yes but have you used it? Because he feels it increases utilisation and I was interested to know if that increase is enough to run only one or two tasks at a time and still have high and efficient use of the 1080/1080ti.

Stephen

? ?
ID: 1952264 · Report as offensive     Reply Quote
Marco Vandebergh

Send message
Joined: 27 Aug 10
Posts: 37
Credit: 11,156,309
RAC: 7,920
Netherlands
Message 1952939 - Posted: 30 Aug 2018, 20:32:43 UTC - in response to Message 1952264.  
Last modified: 30 Aug 2018, 20:47:49 UTC

I am not running it 24/7 due to work and such, i do not leave it unattended, but as long as i can. On regular basis.

It depends on the workunit type what the results are, there are work units that finish in 4 ish minutes when running 1 at a time with -pref_wg_num_per_cu 1,
and 9 ish minutes when running 3 at a time with -pref_wg_num_per_cu 20.

I think it is an improvement., slightly.
GPU load is 99% most of the time, between 97 and 99%.

Interesting to see is that the memory controller usage did go up when using 3 at a time, its 88% to 89 ish
Lowering -pref_wg_num_per_cu 20 also decreases gpu and memory controller usage, so something is happening there.

This is lower when running only 1 or 2 wu's at a time.

I also found, when using cpu_lock and no hyperthreading stuff the GPU more too.


Still trying some settings out though.

I made a screenshot how do i post that?


picture upload
ID: 1952939 · Report as offensive     Reply Quote
Stephen "Heretic" Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 20 Sep 12
Posts: 4621
Credit: 143,154,646
RAC: 245,398
Australia
Message 1953035 - Posted: 31 Aug 2018, 12:54:14 UTC - in response to Message 1952939.  

It depends on the workunit type what the results are, there are work units that finish in 4 ish minutes when running 1 at a time with -pref_wg_num_per_cu 1,
and 9 ish minutes when running 3 at a time with -pref_wg_num_per_cu 20.
I think it is an improvement., slightly.
GPU load is 99% most of the time, between 97 and 99%.
Interesting to see is that the memory controller usage did go up when using 3 at a time, its 88% to 89 ish
Lowering -pref_wg_num_per_cu 20 also decreases gpu and memory controller usage, so something is happening there.
This is lower when running only 1 or 2 wu's at a time.
I also found, when using cpu_lock and no hyperthreading stuff the GPU more too.
Still trying some settings out though.


. .Thanks for that reply, that was enlightening.

Stephen

:)
ID: 1953035 · Report as offensive     Reply Quote
Marco Vandebergh

Send message
Joined: 27 Aug 10
Posts: 37
Credit: 11,156,309
RAC: 7,920
Netherlands
Message 1954282 - Posted: 7 Sep 2018, 19:26:31 UTC - in response to Message 1953035.  
Last modified: 7 Sep 2018, 19:29:40 UTC

Allright, a little update. This will be my final settings for a while. ( or if someone has a suggestion, i can try that )

Now im completing most guppies in 4.25 / 4.40 minutes ish range, when running 2 instances together on the gtx 1080Ti.
In comparison, a single guppi would finish in 2.30 / 2.40 ish minutes range on my system.

The GPU usage is about the same as in my previous post picture above with 3 instances running.
Increasing -pref_wg_num_per_cu 32 any higher does not have a positive effect anymore on my system. Decreasing results in slower completing times of wu. ( on my system )
Afterall i think its best to run just 2.
My Os is win10.

Appconfig.cfg:

<app_config>
<app_version>
<app_name>setiathome_v8</app_name>
<plan_class>opencl_nvidia_SoG</plan_class>
<avg_ncpus>0.4</avg_ncpus>
<ngpus>0.5</ngpus>
<cmdline>-sbs 1024 -spike_fft_thresh 4096 -tune 1 64 1 4 -oclfft_tune_gr 256 -oclfft_tune_lr 16 -oclfft_tune_wg 512 -oclfft_tune_ls 512 -oclfft_tune_bn 64 -oclfft_tune_cw 64 -pref_wg_num_per_cu 32 -period_iterations_num 1 -hp -high_prec_timer -high_perf -tt 1500 -cpu_lock -instances_per_device 2</cmdline>
</app_version>
<app_version>
<app_name>astropulse_v7</app_name>
<avg_ncpus>0.4</avg_ncpus>
<ngpus>0.33</ngpus>
<plan_class>opencl_nvidia_100</plan_class>
<cmdline>-unroll 28 -oclFFT_plan 256 16 256 -ffa_block 12288 -ffa_block_fetch 6144 -tune 1 64 4 1 -tune 2 64 4 1 -hp -cpu_lock -instances_per_device 3</cmdline>
</app_version>
<app_version>
<app_name>astropulse_v7</app_name>
<avg_ncpus>0.4</avg_ncpus>
<ngpus>0.33</ngpus>
<plan_class>cuda_opencl_100</plan_class>
<cmdline>-unroll 28 -oclFFT_plan 256 16 256 -ffa_block 12288 -ffa_block_fetch 6144 -tune 1 64 4 1 -tune 2 64 4 1 -hp -cpu_lock -instances_per_device 3</cmdline>
</app_version>
<app_version>
<app_name>astropulse_v7</app_name>
<avg_ncpus>0.4</avg_ncpus>
<ngpus>0.33</ngpus>
<plan_class>opencl_nvidia_cc1</plan_class>
<cmdline>-unroll 28 -oclFFT_plan 256 16 256 -ffa_block 12288 -ffa_block_fetch 6144 -tune 1 64 4 1 -tune 2 64 4 1 -hp -cpu_lock -instances_per_device 3</cmdline>
</app_version>
<app_version>
<app_name>astropulse_v7</app_name>
<avg_ncpus>0.4</avg_ncpus>
<ngpus>0.33</ngpus>
<plan_class>cuda_opencl_cc1</plan_class>
<cmdline>-unroll 28 -oclFFT_plan 256 16 256 -ffa_block 12288 -ffa_block_fetch 6144 -tune 1 64 4 1 -tune 2 64 4 1 -hp -cpu_lock -instances_per_device 3</cmdline>
</app_version>
</app_config>
ID: 1954282 · Report as offensive     Reply Quote
rob smith Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer moderator
Volunteer tester

Send message
Joined: 7 Mar 03
Posts: 17709
Credit: 402,148,116
RAC: 170,766
United Kingdom
Message 1954617 - Posted: 10 Sep 2018, 8:48:33 UTC

Giys - I've suffered a bit of brain fade here - what is the minimum version of OpenCL required for the SoG application to work?
Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?
ID: 1954617 · Report as offensive     Reply Quote
Stephen "Heretic" Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 20 Sep 12
Posts: 4621
Credit: 143,154,646
RAC: 245,398
Australia
Message 1954619 - Posted: 10 Sep 2018, 9:04:50 UTC - in response to Message 1954617.  

Giys - I've suffered a bit of brain fade here - what is the minimum version of OpenCL required for the SoG application to work?


. . I thought it worked with 1.1 but it certainly works with 1.2 ...

Stephen

? ?
ID: 1954619 · Report as offensive     Reply Quote
Profile Wiggo "Democratic Socialist"
Avatar

Send message
Joined: 24 Jan 00
Posts: 16684
Credit: 225,347,717
RAC: 171,205
Australia
Message 1954620 - Posted: 10 Sep 2018, 9:08:20 UTC - in response to Message 1954619.  

Giys - I've suffered a bit of brain fade here - what is the minimum version of OpenCL required for the SoG application to work?


. . I thought it worked with 1.1 but it certainly works with 1.2 ...

Stephen

? ?
It definitely works on 1.1.

Cheers.
ID: 1954620 · Report as offensive     Reply Quote
rob smith Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer moderator
Volunteer tester

Send message
Joined: 7 Mar 03
Posts: 17709
Credit: 402,148,116
RAC: 170,766
United Kingdom
Message 1954625 - Posted: 10 Sep 2018, 10:40:26 UTC

Ta

Any votes (either way) for 1.0??
Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?
ID: 1954625 · Report as offensive     Reply Quote
Profile Mike Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 17 Feb 01
Posts: 31825
Credit: 74,547,470
RAC: 24,676
Germany
Message 1954637 - Posted: 10 Sep 2018, 12:23:54 UTC - in response to Message 1954625.  

Ta

Any votes (either way) for 1.0??


Yes, it should work with 1.0 as well.
The code itself is 1.0.
It was the last time i asked Raistmer about it.
With each crime and every kindness we birth our future.
ID: 1954637 · Report as offensive     Reply Quote
Previous · 1 . . . 13 · 14 · 15 · 16 · 17 · 18 · 19 . . . 22 · Next

Message boards : Number crunching : SETI@home v8.22 Windows GPU applications support thread


 
©2019 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.