Advanced search

Message boards : Frequently Asked Questions (FAQ) : FAQ - Best configurations for GPUGRID

Author Message
Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 16463 - Posted: 20 Apr 2010 | 15:33:54 UTC
Last modified: 9 Mar 2012 | 11:27:21 UTC

How to report tasks immediately (needed for Boinc 6.12.34 and earlier):

Create, or edit an existing, cc_config.xml file in the Program Data folder for Boinc, using a text editor (Notepad),
- For Vista and Win7 create the file in this folder, C:\ProgramData\BOINC (this folder will likely be hidden, but should open if you type it into Explorer).

Add the following lines:

    <cc_config>
    <options>
    <report_results_immediately>1</report_results_immediately>
    </options>
    </cc_config>



Using Multiple GPUs:
Depending on your driver, you may also need to add this line if you want to use multiple GPUs

    <use_all_gpus>1</use_all_gpus>


Cache Levels (Additional Work Buffer):
For GPUGRID it is better to set a low cache level of say 0.05days (rather than use the default 0.25days).
Open Boinc in Advanced View, Click Advanced, Preferences, under the network usage Tab set the additional work buffer to a low value.
- This will optimise Work Unit turn around and expedite the project. It will also allow you to collect larger bonuses.

HTH
Send message
Joined: 1 Nov 07
Posts: 38
Credit: 6,365,573
RAC: 0
Level
Ser
Scientific publications
watwatwatwatwatwatwatwatwatwat
Message 16588 - Posted: 27 Apr 2010 | 17:55:48 UTC - in response to Message 16463.


Using Multiple GPUs:
You may also need to add this line if you want to use multiple GPUs
    <use_all_gpus>1</use_all_gpus>




I have heard that I should make app_info.xml file for GPUGRID and put these lines in it:
<coproc>
<count>0.5</count>
</coproc>

I don't know how to make it. Do you know? (It needs more than just those three lines.)
____________

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 16832 - Posted: 5 May 2010 | 13:50:33 UTC - in response to Message 16595.
Last modified: 3 Jun 2016 | 21:23:27 UTC

Report tasks immediately on Win XP:
Create the cc_config.xml file here,
C:\Documents and Settings\All Users\Application Data\BOINC\

Report tasks immediately on Vista or win 7:
Create the cc_config.xml file here,
C:\ProgramData\BOINC\

- details in first post -


Also,

You can add this Environmental System Variable to dedicate a CPU core for each GPU (just 1 thread for HT systems),
Variable name: Swan_Sync
Variable Value: 0

XP:
Start, right (alternate) click My Computer, Click Properties - Opens System Properties Window.
Click the Advanced Tab, then Environmental Variables.
Under System Variables, Click New,
For the Variable name type Swan_Sync
For the Variable Value type 0

Vista or Win7:
Start, right (alternate) click Computer, Click Properties - Opens System Window.
Click Advanced System Settings (left side), then Environmental Variables.
Under System Variables, Click New,
For the Variable name type Swan_Sync
For the Variable Value type 0

On Linux you need SWAN_SYNC=1

Paronfesken
Send message
Joined: 16 Jul 09
Posts: 17
Credit: 4,787,632
RAC: 0
Level
Ala
Scientific publications
watwatwatwatwat
Message 16969 - Posted: 11 May 2010 | 20:02:53 UTC

And for ubuntu?

I tried editing the file by typing:

sudo cp /etc/boinc-client/cc-config.xml /etc/boinc-client/backup_cc-config.xml

sudo gedit /etc/boinc-client/cc-config.xml


included the line:

<report_results_immediately>1</report_results_immediately>


in options like this

<options>
<report_results_immediately>1</report_results_immediately>
</options>


then saving and restarting boinc like this

sudo service boinc-client restart


Is this right?


bigtuna
Volunteer moderator
Send message
Joined: 6 May 10
Posts: 80
Credit: 98,784,188
RAC: 0
Level
Thr
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwat
Message 16976 - Posted: 12 May 2010 | 3:03:04 UTC - in response to Message 16969.

That looks about right to me (whatever that is worth).

You can tell if it is working if a project disappears as soon as it is 100% done, rather than the project hanging around with the message "ready to report results".

I just got this working in Ubuntu and I had a heck of a time with file ownership and permissions. I'm accustom to other Linux distros but I'm new to Ubuntu so I'm not familiar with how you level up in Ubuntu.



Paronfesken
Send message
Joined: 16 Jul 09
Posts: 17
Credit: 4,787,632
RAC: 0
Level
Ala
Scientific publications
watwatwatwatwat
Message 16978 - Posted: 12 May 2010 | 7:26:06 UTC

It seams to work and the average work done has raised a bit I think :D

darwincollins
Send message
Joined: 6 Jan 10
Posts: 22
Credit: 105,944,936
RAC: 0
Level
Cys
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwat
Message 17152 - Posted: 19 May 2010 | 6:12:59 UTC - in response to Message 16832.

You said:
For the Variable name type Swan_Sync
For the Variable Value type Swan_Sync=0

perhaps, you meant:
For the Variable name: Swan_Sync
For the Variable Value: 0

that way... it would put Swan_Sync=0 in environment
instead of

Swan_Sync=Swan_Sync=0

right?

Richard Haselgrove
Send message
Joined: 11 Jul 09
Posts: 1576
Credit: 5,517,711,851
RAC: 8,643,215
Level
Tyr
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 17161 - Posted: 19 May 2010 | 18:35:17 UTC - in response to Message 17152.

You would think that (name, value) would be a more natural way of doing it, wouldn't you? I've been running with the whole "Swan_Sync=0" string in the value field ever since host 71413 went online.

This morning, I set the value to be 0: workunit 1490610 (reported 14:13:59) was relatively fast, but within the same range as pervious tasks.

Then I had another thought. Most programmers would use value "1" to activate an option, and "0" for the default setting. So I ran workunit 1491415 (reported 18:30:44) with variable Swan_Sync, value 1.

That run time of 14,588.31 seconds is at least 1,000 seconds faster than any previous run, including my only previous KASHIF_HIVPR_auto_spawn. Might be worth experimenting further.

bigtuna
Volunteer moderator
Send message
Joined: 6 May 10
Posts: 80
Credit: 98,784,188
RAC: 0
Level
Thr
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwat
Message 17162 - Posted: 19 May 2010 | 22:35:27 UTC - in response to Message 17161.
Last modified: 22 Jul 2010 | 21:26:50 UTC

You would think that (name, value) would be a more natural way of doing it, wouldn't you? I've been running with the whole "Swan_Sync=0" string in the value field ever since host 71413 went online.

This morning, I set the value to be 0: workunit 1490610 (reported 14:13:59) was relatively fast, but within the same range as pervious tasks.

Then I had another thought. Most programmers would use value "1" to activate an option, and "0" for the default setting. So I ran workunit 1491415 (reported 18:30:44) with variable Swan_Sync, value 1.

That run time of 14,588.31 seconds is at least 1,000 seconds faster than any previous run, including my only previous KASHIF_HIVPR_auto_spawn. Might be worth experimenting further.


Disclaimer: I'm a total noob and I don't know what I'm talking about...

On multi-core CPUs they start counting at zero so on a quad core you would have core 0, 1, 2, and 3 (instead of 1, 2, 3, 4).

I took swan-sync value "0" to mean [use core number "0"] (meaning core 1 to non-geeks). If I'm right swan-sync "1" would use core 2 etc.

Does anyone run a CPU monitor? If so you might see a different core being used as you cycle the numbers??

EDIT:

Evidently this is total nonsense (my understanding was incorrect).

See this post for an explanation of Swan_Sync:

http://www.gpugrid.net/forum_thread.php?id=2227&nowrap=true#18136

Richard Haselgrove
Send message
Joined: 11 Jul 09
Posts: 1576
Credit: 5,517,711,851
RAC: 8,643,215
Level
Tyr
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 17163 - Posted: 19 May 2010 | 23:30:23 UTC - in response to Message 17162.

Value 1 on a TONI_HERGunb1 (workunit 1493006, same host as before) is back to the same sort of time as before. I'm about tested out for tonight: maybe tomorrow I'll try it with an empty value string, or even no variable at all, and see if it slows down - maybe just having anything at all there is good enough.

Richard Haselgrove
Send message
Joined: 11 Jul 09
Posts: 1576
Credit: 5,517,711,851
RAC: 8,643,215
Level
Tyr
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 17172 - Posted: 20 May 2010 | 11:33:02 UTC - in response to Message 17163.
Last modified: 20 May 2010 | 11:53:11 UTC

Ran workunit 1494710 with no environment variable at all - slightly slower, but nothing like the 20% - 30% envisaged by the original announcement.

From the text of that message, I think it's clear that the expected value is just 0

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 17175 - Posted: 20 May 2010 | 13:44:05 UTC - in response to Message 17172.
Last modified: 14 Oct 2011 | 15:46:09 UTC

For Ubuntu:

From a Terminal goto the folder containing the cc_config.xml file by typing,

    cd /etc/boinc-client/



Then backup the file by typing,

    sudo cp /etc/boinc-client/cc_config.xml /etc/boinc-client/backup_cc_config.xml



Then edit the file by typing,

    sudo gedit /etc/boinc-client/cc_config.xml



In options included the line:

    <options>
    <report_results_immediately>1</report_results_immediately>
    </options>



then save and restart boinc,

    sudo service boinc-client restart



---

I don’t think that swan_sync works for 6.03 tasks, (compiled for CUDA 2.2). It is there for CUDA 3 or above (the forthcoming 3.1 that might be 11% faster than CUDA 3.0, which is 30-35% slower than CUDA2.2 which does not work for Fermis, so this might even things out).
So it is most notably for Fermi, but note that it requires the cruncher to leave one CPU thread free per GPU, to fully benefit from the Environmental Variable.
If you wanted to use four GPU's a single CPU core is not enough (for all 4 GPUs), you need a Quad core CPU, or better.

I think the variable may work for the more recent 6.72 WU's but to a lesser extent. Either way, it should do soon (if/when these are compiled for CUDA 3.0, or more likely 3.1).
-
Swan_sync now works for the 6.12app and with non-Fermi cards as well as the 6.13app for Fermi cards or some pre-Fermi cards.

Richard Haselgrove
Send message
Joined: 11 Jul 09
Posts: 1576
Credit: 5,517,711,851
RAC: 8,643,215
Level
Tyr
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 17177 - Posted: 20 May 2010 | 14:03:30 UTC - in response to Message 17175.

All my experiments with SWAN_SYNC in this thread relate to host 71413, which as you can see is an eight-core host (that's eight real cores, not HT semi-cores). I'm running seven BOINC CPU tasks plus one GPUGrid task on the single GTX 470 GPU. Obviously, Windows doesn't exactly run 'one task per core' - the thermal stresses would be horrendous if seven cores were generating maximum heat and the eighth was idle - but there should be plenty of spare cycles for the syncing.

If you look back at the announcement link from last time, you'll see that it is in the "GPUGRID and Fermi" thread. Also, it's clear from that announcement that the 20%-30% speedup I mentioned refers to a single application - by implication, v6.73 - with two modes of operation. That's what I'm failing to see.

I'm not sure where this idea come from, of a physical core ID being specified for the sync process - that goes against everything I've ever learned about multi-tasking operating systems - and it isn't supported by any post I can find from the project developers. But perhaps as you say we should wait for technical clarification.

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 17302 - Posted: 25 May 2010 | 18:22:13 UTC - in response to Message 17177.
Last modified: 6 Dec 2010 | 23:09:26 UTC

It does not use one single CPU core, and on multi-GPU systems the CPU usage is spread out over all CPU cores, and controlled by the CPU, but overall 4 CPU cores will be used for a system with 4 GPU's.

L2B
Send message
Joined: 2 Jul 10
Posts: 3
Credit: 7,394,761
RAC: 0
Level
Ser
Scientific publications
watwatwatwatwatwatwatwatwat
Message 18201 - Posted: 29 Jul 2010 | 5:36:49 UTC

L2B here...

Hello, I'm pretty new here, have been running Grid tasks for a few weeks and am looking for some feedback

I'm not currently getting any tasks, any ideas why not?

Currently running a pair of 465s, a i7-930 and the only thing overclocked are the fan settings on the GPUs!

Thanks

L2B
Send message
Joined: 2 Jul 10
Posts: 3
Credit: 7,394,761
RAC: 0
Level
Ser
Scientific publications
watwatwatwatwatwatwatwatwat
Message 18202 - Posted: 29 Jul 2010 | 5:58:20 UTC

Ok...I'm back running Grid tasks...Brain fade on my part!

See previous post for my configuration.

It appears that I should be able to complete a ...dynamics 6.11 (cuda31) task in about 3.5 hours.

Does that seem sensible, given other's experience?

Thanks

bigtuna
Volunteer moderator
Send message
Joined: 6 May 10
Posts: 80
Credit: 98,784,188
RAC: 0
Level
Thr
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwat
Message 18206 - Posted: 29 Jul 2010 | 18:58:53 UTC - in response to Message 18202.

Ok...I'm back running Grid tasks...Brain fade on my part!

See previous post for my configuration.

It appears that I should be able to complete a ...dynamics 6.11 (cuda31) task in about 3.5 hours.

Does that seem sensible, given other's experience?

Thanks

The tasks are different "sizes" so some take longer to run than others. Points awarded are typically higher for longer running tasks.

3.5 hours for a 6800 point work unit seems about right. Some work units will take about twice as long but they should also be worth about twice as much.

Michael E. Malis
Send message
Joined: 15 May 10
Posts: 7
Credit: 11,913,040
RAC: 0
Level
Pro
Scientific publications
watwatwat
Message 18259 - Posted: 6 Aug 2010 | 4:48:42 UTC

I'm not sure what the specific cause of it is but DO NOT USE THE SWAN_SYNC WITH PHYS-X DEDICATED TO THE CPU IF YOU ARE RUNNING VISTA 64-BIT. It causes Windows Explorer to continuously stop working and then restarts Explorer over and over in an infinite loop and is extremely difficult to maneuver to change the settings back because you only have about 8 seconds in-between Explorer crashes.

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 19851 - Posted: 10 Dec 2010 | 12:11:50 UTC - in response to Message 18259.
Last modified: 21 Dec 2010 | 10:28:25 UTC

The more recent NVidia drivers are causing many GPUs to reduce their clock rates. So it is a good idea to make sure your cards Power management is set to Prefer Maximum Performance (NVidia control panel, Manage 3D settings, Power management mode).

I recommend that people crunching on Fermi cards control their fan speed using a tool such as EVGA Precision (now available without email registration). Typically these cards will run at up to 91 deg C. At such temperatures the task success rate drops. It is best to keep temperatures under 80deg C and under 70deg C if possible. Doing so will increase task success rates and the cards longevity.

kts
Send message
Joined: 4 Nov 10
Posts: 21
Credit: 25,973,574
RAC: 0
Level
Val
Scientific publications
watwatwatwatwatwatwatwat
Message 20113 - Posted: 5 Jan 2011 | 1:32:14 UTC - in response to Message 16590.

To run GPUGrid tasks on a Fermi GPU find the app_info.xml file and include the following,

<app_info>
<coproc>
<type>CUDA</type>
<count>0.5</count>
</coproc>
</app_info>


I have added this into my W7 \program data\BOINC folder, but any explanation what it means/does to BOINC in general and GPUGRID specifically? Impact on other gpu projects? Other <coproc> flags documented?

Thanks

dasfiend
Send message
Joined: 1 Apr 09
Posts: 2
Credit: 17,884,639
RAC: 0
Level
Pro
Scientific publications
watwatwatwatwatwatwatwat
Message 20160 - Posted: 13 Jan 2011 | 1:41:06 UTC

I just installed a new gtx 570 (evga factory superclocked edition) and am trying to get the most from it.

After putting an app_info.xml as described above in D:\boinc\data\projects\www.gpugrid.net (D:\boinc\data is my data dir)

I am no longer getting any work units and
I get the following strange entries in my log (see code snippet below).

What am I doing wrong?




1/12/2011 7:35:35 PM NVIDIA GPU 0: GeForce GTX 570 (driver version 26309, CUDA version 3020, compute capability 2.0, 1248MB, 1530 GFLOPS peak)
1/12/2011 7:35:35 PM GPUGRID Found app_info.xml; using anonymous platform
1/12/2011 7:35:35 PM GPUGRID [error] State file error: missing application acemd2
1/12/2011 7:35:35 PM GPUGRID [error] Can't handle workunit in state file
1/12/2011 7:35:35 PM GPUGRID [error] State file error: missing application acemd2
1/12/2011 7:35:35 PM GPUGRID [error] Can't handle workunit in state file
1/12/2011 7:35:35 PM GPUGRID [error] State file error: missing task 8-IBUCH_smdwt_EGFR_110112-0-2-RND6116
1/12/2011 7:35:35 PM GPUGRID [error] Can't link task 8-IBUCH_smdwt_EGFR_110112-0-2-RND6116_1 in state file
1/12/2011 7:35:35 PM GPUGRID [error] State file error: missing task 83-KASHIF_HIVPR_n1_unbound_so_ba2-61-100-RND8936
1/12/2011 7:35:35 PM GPUGRID [error] Can't link task 83-KASHIF_HIVPR_n1_unbound_so_ba2-61-100-RND8936_0 in state file
1/12/2011 7:35:35 PM GPUGRID [error] State file error: result 8-IBUCH_smdwt_EGFR_110112-0-2-RND6116_1 not found for task
1/12/2011 7:35:35 PM rosetta@home URL http://boinc.bakerlab.org/rosetta/; Computer ID 1401747; resource share 100
1/12/2011 7:35:35 PM GPUGRID URL http://www.gpugrid.net/; Computer ID 91162; resource share 100
1/12/2011 7:35:35 PM No general preferences found - using BOINC defaults
1/12/2011 7:35:35 PM Reading preferences override file
1/12/2011 7:35:35 PM Preferences:
1/12/2011 7:35:35 PM max memory usage when active: 8190.26MB
1/12/2011 7:35:35 PM max memory usage when idle: 14742.47MB
1/12/2011 7:35:35 PM max disk usage: 10.00GB
1/12/2011 7:35:35 PM don't use GPU while active
1/12/2011 7:35:35 PM suspend work if non-BOINC CPU load exceeds 25 %
1/12/2011 7:35:35 PM (to change preferences, visit the web site of an attached project, or select Preferences in the Manager)
1/12/2011 7:35:35 PM Not using a proxy

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 20163 - Posted: 13 Jan 2011 | 14:03:25 UTC - in response to Message 20160.

Close Boinc and delete the app_info.xml file. You would have needed to include additional details about the apps to use.

Running multiple tasks on the one GPU is not recommended; it only speeds up running some task types, and these are being phased out.

Content yourself with using the following optimizations:

Creating the swan_sync environmental variable and setting it to zero (0).
Freeing up one CPU core/thread per GPU; use 99% CPU's for one GPU.
Configure the cc_config.xml file to report tasks immediately.
Keep a low Boinc cache.

Increasing the CPU and GPU speeds modestly can increase performance, but avoid overclocking to the point of getting failures as this will reduce overall turnover. Overclocking the GPU a lot can actually increase run times, so make sure your OC is working by comparing completion times of same task types WU's with the same credit. It is usually not worth increasing the voltage, as this will increase power usage, heat, noise and reduce the MTF. Increasing fan speed is always recommended for GPU crunching.

Profile [AF>EDLS] Polynesia
Avatar
Send message
Joined: 3 Apr 09
Posts: 11
Credit: 5,336,576
RAC: 0
Level
Ser
Scientific publications
watwatwatwatwatwatwatwatwat
Message 20168 - Posted: 15 Jan 2011 | 21:53:25 UTC - in response to Message 20163.

must put 99% in the preferences of boinc manager to release a CPU, right?

because I tried with swan_sync and I do not see much difference ....!

(you have rebooted after changing the variable swan_sync?)
____________
(boinc since 1/04/09), Alliance Francophone
GA-P55A-UD5, i7 860 2.8 Ghz, Win 7 64 bits, 8 g DDR3 de ram, 2 To DD + 750 ext, boitier Haf 922 + 2 ventilo 120 noctua, ventirad noctua NH-U12P SE2, GTX 470 zotac amp edition, écran 25"

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 20169 - Posted: 16 Jan 2011 | 18:25:19 UTC - in response to Message 20168.
Last modified: 11 May 2011 | 20:46:28 UTC

Yes. You need to use SWAN_SYNC and free a CPU core/thread as well.

How to stop tasks switching:
Keep a low cache and set set the processor usage to switch between applications every 999sec.
In theory this might affect the running of a long task when a shorter task then downloads, or when you crunch for more than 1 GPU project.

On Vista and Win7 change the NVidia power scheme from adaptive to max performance.

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 21283 - Posted: 1 Jun 2011 | 9:35:11 UTC - in response to Message 20169.
Last modified: 23 Feb 2014 | 18:01:46 UTC

There are four 'processor usage' local Boinc settings that could cause the GPU to suspend and resume frequently. Boinc Manager (Advanced view), Tools, Computing Preferences, Processor Usage tab,
Computing allowed,

While computer is in use (when not ticked)
Use GPU while computer is in use (when not Ticked)
While processor usage is less than (anything other than zero)
Use at most % CPU time (anything other than 100%)

In particular,
- less than 100% CPU time is bad
Also,
- Using TThrottle can cause tasks to stop and start hundreds of times
- if a Kepler GPU with boost gets too hot / loud / expensive: reduce the power target, as this makes the chip run more efficiently
See this and subsequent posts.

SLI is not recommended for GPUGrid.

Only one task can be run on one card, not one task across 3 cards simultaneously (Might change in the future).

While you could run 2 tasks on one card, by configuring an app_info.xml or app_config.xml file, this is generally not recommended; tasks tend to run at >90% GPU Utilization here, so there is no benefit. You could also run out of GDDR memory.

If you experience GPU downclocking open NVidia control panel/Xserver and
change the PowerMizer from "Adaptive" mode to Max Performance if it's available (not always there on XP, but usually is on Vista, Win7 and some Linux systems). Ubuntu 12, 11.04 and 11.10 versions (and others) do not save your performance selection, so you will need to reselect it each time you restart! This often prevents the GPU from down-clocking. If it still downclocks you should try restarting your system, check your temperatures, configure fan controlling software and try reducing your GDDR frequency if need be.

Linux - GPU not found:
One off fix (restart the Boinc client),

    sudo /etc/init.d/boinc-client restart

Permanent fix (prevent Boinc starting before the drivers are loaded),

    sudo update-rc.d -f boinc-client remove
    sudo update-rc.d boinc-client defaults 99

or,

    sudo gedit /etc/init.d/boinc-client

Ref, add sleep time

See this Manual Fan Control Guide for setting the fan speed under Gentoo Linux. You can also follow these steps to setup fan control for Ubuntu 11.10 (and similar). Restart to apply the setting.

It is recommended to Not use a screen saver at all. Better to manually turn the monitor off. Behaviour varies from driver to driver with downclocking and crashing common.
See Jacob Klein's post here on the Win7+DVI+295/296 driver issue.

Disabling Aero and visual effects (or not installing) increases GPU utilization (Win7 and Vista).

Remember to choose short or long tasks according to your GPU's performance and the amount of time it runs GPUGrid tasks.
You can do this from Your account, GPUGRID preferences
For those crunching Long tasks only, I suggest selecting to accepting normal tasks if no long tasks are available,

    If no work for selected applications is available, accept work from other applications? yes


It is not recommended to run multiple GPU projects on the same system. However if you decide to do this it is recommended that you keep a low cache and configure Boinc Manager to "switch between applications every" to a high number (999)minutes. If you have more than one GPU it's better to specify which GPU to use for which project.

Remote desktop - Don't use it! RDP uses a default windows video driver and causes problems with work units.

Multiple GPU setups.
It is best to use GPU's of similar Compute Capability. Mixing CC1.3 and CC2.0 cards, for example, might result in some errors (mostly for the CC1.3 GPU). Fermi's only differ by being CC2.0 or CC2.1, but even this could be a problem depending on drivers.

If you're using Ubuntu 12.x and are missing libs open a terminal try,
sudo apt-get install openssh-server+ gdebi+ libwxgtk2.8-0+ libXss1+ freeglut3+ gnome-applets+ cpufrequtils+ ia32-libs+
(Thanks to Steve and Zombie for this)

Power Settings - If Sleep is enabled computation will stop after the period you set to sleep after. Also note that PCIE and Network power settings in Linux, Vista (especially), W7 and potentially W8 might cause performance/communication issues. I recommend setting these power options to full power (energy ways there is little to be gained from a built in Network card going to sleep).

Adding <use_all_gpus>1</use_all_gpus> to a cc_config.xml file is recommended for systems with more than one usable GPU.
Although the GPUGrid server asks for the tasks to be returned immediately, earlier clients do not respond. So you might want to use <report_tasks_immediately>1</report_tasks_immediately>
Not needed for Boinc versions 7.0.2x onwards.

When installing new drivers, exit Boinc (and close running tasks), install the driver (clean install preferred), and then restart the system.

It's better to use the latest recommended version of Boinc, but if you are a tester you can download the latest version from http://boinc.berkeley.edu/dl/?C=M;O=D
____________
FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Message boards : Frequently Asked Questions (FAQ) : FAQ - Best configurations for GPUGRID

//