Need SETI@home GPU opt app advice

[EmSti]

This could be interesting....I am completing cuda work units now that I have lunitics code in place. But it is taking longer to download a workunit than it takes a gtx 680 engine to complete it. First 3 wus took 2 minutes, 2 minutes and 6 minutes to complete. 2 tasks are downloading at a time, some suspending off and on, and taking over 3 minutes.

Will let it go for awhile and see if seti starts sending any faster.

[Slicker]

Okay, I looked at the top hosts. AMD cards are *WAY* down the list. So it looks like there is no point in using AMD cards there.

So I am refining my question. Best settings for nVidia cards listed in the previous post?

That's because SETI uses OpenCL for AMD. OpenCL performance can be compared to the stuff that comes out the southbound end of a northbound mule. And yet, AMD has dropped CAL support in the latest drivers (or at least the beta version of them). Also, remember that OpenCL is 90% based on CUDA and, whether or not people really want to admit it, the language was developed for the nVidia hardware whereas AMD has had to develop their hardware to work with the nVidia preferred software. And did I mention that nVidia provided development support to BOINC/SETI whereas AMD did not?

Then again, nVidia still supports CUDA and OpenCL on an 8400GS. AMD has dropped support for HD 38xx and now HD 48xxx GPUS.

My suggestion is that if you have AMD stock, sell.

[Duke of Buckingham]

Very nice I installed the application and now I am running SETI on ATI, thanks Z.

[EmSti]

Adding to collective knowledge the little bit of info I have gained based on my system and SETI.

My setup 1 7970 and 1 gtx 690 (2 gtx 680 engines - added to help with search hits) with i7-2600k, running lunitics with a patch for ati/nvidia mixed setups.

For the 680 engines running 3 multiBeeam (labeled as setihome_enhanced application) wus per GPU (6 total). Better than 2, I haven't tried 4 since I am getting 96 to 98% GPU utilization.

For 7970, I got a lot of restarts on the ATI multibeam wus until they errored out. Found the error and a patch for lunatic, but Seti hasn't resent me any ati MultiBeam wus since the patch, so I don't know if I added the files properly or if it worked. Refer to
http://setiathome.berkeley.edu/forum...d.php?id=66812
Look at the last post for download. If someone has a newer install of lunatic than v 0.40, please share. Since this isn't working yet, it could be a patch for a newer version or I have it in wrong, or I have to rest to get Multibeam again. A nwer version may already include the patch. Date of files in the patch 2/3/2012.

For astropulse on 7970, I have made some progress on completion times.
Original Lunitics config: <cmdline>-instances_per_device 1 -unroll 4 -ffa_block 2048 -ffa_block_fetch 1024 -sbs 128 -v505</cmdline>
Original times: about 33 to 35 minutes per wu
Gpu usage: Don't remember exactly, but it bounced around 50%.

Change 1: <cmdline>-instances_per_device 1 -unroll 16 -ffa_block 8192 -ffa_block_fetch 4096 -sbs 512 -hp -no_cpu_lock </cmdline>
Change 1 times: about 26 minutes to 28 minutes per wu
unroll 16 - picked that value because it is 1/2 of the cumputation units of 7970 (see link below)
Rest of the values are guesses based on some light reading, there could be better values. I would love corrections.
Change 1 GPU: GPU bounces between 67 to 80%

Change 2: <cmdline>-instances_per_device 2 -unroll 16 -ffa_block 8192 -ffa_block_fetch 4096 -sbs 512 -hp -no_cpu_lock </cmdline>
<coproc>
<type>ATI</type>
<count>0.5</count>
</coproc>
Change 2 times: About 2 wus every 41 to 42 minutes or 1 wu every 21 minutes. Sticking with the params for now.
Change 2 GPU: nearly 100%, bounced between 88 to 96% (edited to correct values)

Link to param definitions.
http://setiathome.berkeley.edu/forum...sort=5#1210310

Biggest problem is getting astropulse downloaded fast enough to get 2 wus to run at the same time. Faster to crunch them than to download them so far. Starts out only allowing a small number, hopefully client will tune the times and get more at once.

[EmSti]

I have learned of a feature in boinctasks that may be helpful. Since Seti downloads seem to constantly stop and retry later (often hours), I went looking for a way to do retry all on the downloads automatically. Looks like BoincTasks added a tunable feature to do updates and retries. I think it only works on the localhost or boinc tasks, someone with more than one machine will have to test to multiple clients.

Still experimenting but it seems to help speed up the seti retries. I added the following config.xml to the exe dir for BonicTasks and restarted bonic taks. The downloads and uploads parts are commented out, using the auto. You can see if it took by looking the in bonictasks log (from the menu list Show -> log).

<config>
<refresh>
<!-- <downloads> 3600 </downloads>
<uploads> 3600 </uploads> -->
<auto> 600 </auto>
<refresh> 900 </refresh>
</refresh>
</config>

<!--

Make sure you open this file in a simple text editor, like notepad.exe only.
This file has to be stored in ascii format.

Remove any items that you don't need.
Make sure there is only one <config></config> block!

Next store this file next to the BoincTasks.exe and restart BoincTasks.

options block:

no_cleanup 1 Will not cleanup any data, like e.g. - 13% -> 0%, 123% -> 100% or remaining time -1d -> 0
0 Will cleanup the data.

checkpoint block:

Generate a warning color on overdue checkpointing.

refresh block:

downloads retry all downloads on the localhost every xx seconds.
uploads retry all uploads on the localhost every xx seconds.
auto will disable downloads, uploads. The refresh time will be auto regulated. Uses the value for the maximum retry time.
refresh refresh on the localhost every xx seconds. = Report all completed tasks.

-->

Edit: The comment section at the end of the config.xml is from the example found in \Program Files\eFMer\BoincTasks\examples\config.xml