Number of Seti tasks on a Quadro 3800

Message boards : Number crunching : Number of Seti tasks on a Quadro 3800
Message board moderation

To post messages, you must log in.

AuthorMessage
PhilG57

Send message
Joined: 18 Nov 03
Posts: 17
Credit: 17,538,280
RAC: 11
United States
Message 1753283 - Posted: 2 Jan 2016, 18:31:32 UTC

I have a computer with a NVIDIA Quadro FX 3800 with 1024MB of memory. According to the web, this card has 192 CUDA cores. When running BOINC SETI, how many tasks should show up as running on this card? My system show only 1. Thanks.
ID: 1753283 · Report as offensive
Profile Bill G Special Project $75 donor
Avatar

Send message
Joined: 1 Jun 01
Posts: 1282
Credit: 187,688,550
RAC: 182
United States
Message 1753319 - Posted: 2 Jan 2016, 20:27:42 UTC - in response to Message 1753283.  

That is all that would show up by default. That is without modifications.

SETI@home classic workunits 4,019
SETI@home classic CPU time 34,348 hours
ID: 1753319 · Report as offensive
Profile Michael W.F. Miles
Avatar

Send message
Joined: 24 Mar 07
Posts: 268
Credit: 34,410,870
RAC: 0
Canada
Message 1753323 - Posted: 2 Jan 2016, 20:32:00 UTC - in response to Message 1753283.  

Hey bud

Your computer is hidden so can't see if you run Anonymous Tasks or not.
If you are then you will need to edit your app_info.xml file in the C:\ProgramData\BOINC\projects\setiathome.berkeley.edu folder

Set folder options- View and check show hidden files if you have not already
What you need to change in app_info.xml is here

Look below, this just a sample of my app_info.xml entry
Where you see <count>1</count> change the 1 to .5 if you wish to run 2 tasks at once. It would be .3 for 3 tasks.

Your memory on the card is what limits you. For 1 gig 2 tasks would be good. 3 or more you would need more ram on the card.

<app_version>
<app_name>astropulse_v7</app_name>
<version_num>710</version_num>
<platform>windows_x86_64</platform>
<avg_ncpus>0.04</avg_ncpus>
<max_ncpus>0.2</max_ncpus>
<plan_class>opencl_nvidia_100</plan_class>
<cmdline></cmdline>
<coproc>
<type>CUDA</type>
<count>1</count>
</coproc>
<file_ref>
<file_name>AP7_win_x86_SSE2_OpenCL_NV_r2887.exe</file_name>
<main_program/>
</file_ref>
<file_ref>
<file_name>libfftw3f-3-3-4_x86.dll</file_name>
</file_ref>
<file_ref>
<file_name>AstroPulse_Kernels_r2887.cl</file_name>
</file_ref>
<file_ref>
<file_name>ap_cmdline_win_x86_SSE2_OpenCL_NV.txt</file_name>
<open_name>ap_cmdline.txt</open_name>
</file_ref>
</app_version>



Hope this helps

Michael Miles
ID: 1753323 · 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 1753343 - Posted: 2 Jan 2016, 21:17:23 UTC

The issue with running more than 2 or three tasks has on your GPU has NOTHING to do with the amount of memory, but a lot to do with the way the GPU handles its processing units.
Each GPU/CPU/motherboard combination has a sweet-spot for performance which can only be found by a bit of trial and error. You've been running one task on the GPU for a bit, so you have an idea of the RAC that system is delivering, so now increase it to 2 tasks by decreasing the <count> to 0.5 (two tasks at a time on the GPU), leave it there for a couple of weeks (APs are rare, so you have to wait for them to be delivered), and then try it with 0.3 (three AP tasks at a time on the GPU). It is very rare that an Nvidia GPU sweet-spot will be more than 3 tasks per GPU.
Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?
ID: 1753343 · Report as offensive
PhilG57

Send message
Joined: 18 Nov 03
Posts: 17
Credit: 17,538,280
RAC: 11
United States
Message 1753555 - Posted: 3 Jan 2016, 12:21:11 UTC

Thanks for the responses. There is no 'app_info' file on my system so I can check that out. No 'cc_info' file either for that matter which I've seen mentioned in other posts.

BTW, it's an H-P workstation, 2 Xeon 4 way processors, and lots of memory. By enabling virtualization for the processors, I now see 16 Seti work units churning away. That caused me to wonder if I was getting max throughput from my NVIDIA card.

*Question*: I do not process Astro Pulse work units; do the suggestions for changing 'app_info' also apply to the basic Seti work units?

Thanks.
ID: 1753555 · Report as offensive
Darth Beaver Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Avatar

Send message
Joined: 20 Aug 99
Posts: 6728
Credit: 21,443,075
RAC: 3
Australia
Message 1753561 - Posted: 3 Jan 2016, 12:59:00 UTC - in response to Message 1753555.  

Thanks for the responses. There is no 'app_info' file on my system so I can check that out. No 'cc_info' file either for that matter which I've seen mentioned in other posts.

BTW, it's an H-P workstation, 2 Xeon 4 way processors, and lots of memory. By enabling virtualization for the processors, I now see 16 Seti work units churning away. That caused me to wonder if I was getting max throughput from my NVIDIA card.

*Question*: I do not process Astro Pulse work units; do the suggestions for changing 'app_info' also apply to the basic Seti work units?

Thanks.


The "cc_onfig "file you have to make yourself and can be made buy opening note pad and then saving the blank file as cc_config.xml

The appinf.xml is part of a system called Lunatics and it will allow you to do more than one unit on the GPU but does nothing for CPU units

Turning Virtualisation on the MoBo can allow you to have more CPU's it's also called Hyper threading and doubles the cores you have buy way of virtualisation you don't actually have any more physical cores but there are virtual cores.

It slows down the speed of the individual cores however because you have more cores you will do more work about 30% more so expect the times to take a little longer to do now you have Virtual cores

This is a link to Lunatics at Crunchers Anonymous where you can download it but I would wait for a while till the new version is ready and can allow you to do the new Seti V8 new apps otherwise you will get errors in the log file as Lunatics is not ready for the new V8 version of apps .Your running stock so you will be able to do V8 units without the errors as you don't need to change anything when running stock Seti apps in Bionic . Lunatics is a bit different so just wait for the new version in a few weeks to come out

http://www.arkayn.us/forum/
ID: 1753561 · Report as offensive
Richard Haselgrove Project Donor
Volunteer tester

Send message
Joined: 4 Jul 99
Posts: 14649
Credit: 200,643,578
RAC: 874
United Kingdom
Message 1753563 - Posted: 3 Jan 2016, 13:08:27 UTC - in response to Message 1753561.  

A better way to create a cc_config.xml file (assuming a current version of BOINC) is to open the Event Log options pane (Ctrl-Shift-F) and make a minor change - cpu_sched and sched_op_debug are always useful - and click 'OK'. That will write a complete, properly formatted, cc_config template with default values filled in. Much the quickest and safest way to start.

BUT - it has nothing to do with the question about running more than one task on a GPU. For that, you can create a file called app_config.xml - the instructions are in Application configuration
ID: 1753563 · Report as offensive
Captiosus

Send message
Joined: 28 Mar 08
Posts: 2
Credit: 1,144,797
RAC: 0
United States
Message 1753825 - Posted: 4 Jan 2016, 10:17:05 UTC

Erm, I would advise against running multiple GPU units on that Quadro. That card uses the GT200L core, which is a variant of the same core found on the GTX 260, 275, 280, etc.

*Deeper hardware level stuff below*
The problem with this is that the architecture used on the GT200 series chips is NOT well suited for multi-tasking due to the way it schedules work. The GPU could only work on a single subtask or kernel at a time across the entire chip. If that kernel wasn't wide enough (that is, calling for all available resources) the unused resources would go idle until the next kernel began execution.
So even if you did set it up to run 2 tasks at the same time, the way the scheduling works on those older chips would mean that there would be at best a slight performance increase, and more likely, performance regression as the two tasks fight each other over scheduling slots.
Its the reason why Nvidia cooked up Fermi, as Fermi was designed and built from the ground up to do precisely this. If a kernel isnt wide enough to use the entire GPU, it grabs a chunk of another one and uses the remaining cores to start chewing through it instead.
See: http://www.anandtech.com/show/2977/nvidia-s-geforce-gtx-480-and-gtx-470-6-months-late-was-it-worth-the-wait-/3
*Deeper hardware level stuff ends*

tl;dr Your GPU is simply too old and outdated, does not have multi-threading in mind, and making it run 2-3 workunits would simply be wasting time.
ID: 1753825 · Report as offensive
PhilG57

Send message
Joined: 18 Nov 03
Posts: 17
Credit: 17,538,280
RAC: 11
United States
Message 1753839 - Posted: 4 Jan 2016, 12:32:42 UTC
Last modified: 4 Jan 2016, 12:35:27 UTC

Yes, thanks again to all for the help. I'm comfortable running just one Seti (non-AstroPulse) work unit on the GPU. The reason for the post was to ensure I was not missing some opportunity to use all 192, or so, cores.

BTW, I have noticed that CUDA work units seem to arrive on my computer in bunches with periods of no CUDA tasks available for several hours. Is this normal?

Thanks again.
ID: 1753839 · 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 1753845 - Posted: 4 Jan 2016, 13:43:21 UTC

The sporadic nature of work for GPUs is fairly normal, it depends on the "tape" being split, if the tape results in a lot of "VLARs" these are not sent to GPUs because they give the lower performance Nvidia ones a bit of a headache.

btw - running one task will use all 192 cores at once as they are used in parallel.
Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?
ID: 1753845 · Report as offensive
Profile Jimbocous Project Donor
Volunteer tester
Avatar

Send message
Joined: 1 Apr 13
Posts: 1849
Credit: 268,616,081
RAC: 1,349
United States
Message 1753947 - Posted: 4 Jan 2016, 21:32:30 UTC - in response to Message 1753825.  

Erm, I would advise against running multiple GPU units on that Quadro. That card uses the GT200L core, which is a variant of the same core found on the GTX 260, 275, 280, etc.

+1

I've got a 3700 and a 3800, and tested each both ways. Definitely 2 jobs per is a no-no in terms of performance. Even under-performed relative to my GT610s ...
ID: 1753947 · Report as offensive
Captiosus

Send message
Joined: 28 Mar 08
Posts: 2
Credit: 1,144,797
RAC: 0
United States
Message 1753973 - Posted: 4 Jan 2016, 23:01:35 UTC - in response to Message 1753947.  

Erm, I would advise against running multiple GPU units on that Quadro. That card uses the GT200L core, which is a variant of the same core found on the GTX 260, 275, 280, etc.

+1

I've got a 3700 and a 3800, and tested each both ways. Definitely 2 jobs per is a no-no in terms of performance. Even under-performed relative to my GT610s ...

See, if I was in his situation I would consider putting some thought into getting a GTX 600 series GPU or something GCN based on the AMD side. I have a GTX 660 that does well running two units, as running a single unit leaves about 40% of it idle.
And actually, is SETI@Home double precision floating point heavy? Cause if it is he'd be better off either getting a first gen Fermi card, a first gen Titan, or going to the red team.
ID: 1753973 · Report as offensive
Richard Haselgrove Project Donor
Volunteer tester

Send message
Joined: 4 Jul 99
Posts: 14649
Credit: 200,643,578
RAC: 874
United Kingdom
Message 1753974 - Posted: 4 Jan 2016, 23:03:12 UTC - in response to Message 1753973.  

And actually, is SETI@Home double precision floating point heavy? Cause if it is he'd be better off either getting a first gen Fermi card, a first gen Titan, or going to the red team.

No, it isn't. Almost exclusively single precision.
ID: 1753974 · Report as offensive
PhilG57

Send message
Joined: 18 Nov 03
Posts: 17
Credit: 17,538,280
RAC: 11
United States
Message 1754358 - Posted: 6 Jan 2016, 15:30:30 UTC

'Morning everyone. It is nice to have confirmed all the cores in the FX 3800 are being used and that its running only one work unit at a time is probably the best strategy. I appreciate the suggestions on a new graphics card but I'm retired, living on a fixed income, so purchasing another, better, card is out of the question. It is reassuring to know I'm getting the best performance I can out of this system. Many thanks to all...
ID: 1754358 · Report as offensive

Message boards : Number crunching : Number of Seti tasks on a Quadro 3800


 
©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.