Welcome to MilkyWay@home

Using the MW ATi application, Q & A

Message boards : Number crunching : Using the MW ATi application, Q & A
Message board moderation

To post messages, you must log in.

Previous · 1 · 2 · 3 · 4 · 5 · Next

AuthorMessage
Profile ritterm
Avatar

Send message
Joined: 16 Jun 08
Posts: 93
Credit: 366,882,323
RAC: 0
Message 28798 - Posted: 3 Aug 2009, 12:39:28 UTC - in response to Message 28771.  

There is a way around it, but it does mean that your CPU won't get any work.

Change to <avg_ncpus>1</avg_ncpus> or however many threads your system thinks it has (so if it's a P4 with hyperthreading, 2). That means only one WU will say it is working at a time.

The AGP cards are a bit flaky - there's nothing we can do.

This seems to work perfectly for what I need! For the past 12 hours, this host has been crunching one WU at a time, one right after another. And, they run in about half the indicated CPU time, as well.

The response is a little sluggish, but, fortunately for me, this host is only for BOINC and as a platform to learn a bit about Linux.

Thanks again, to everybody, for helping me understand a little better how this works! :-)

MarkR



ID: 28798 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Sharon

Send message
Joined: 11 Apr 08
Posts: 10
Credit: 17,339,474
RAC: 0
Message 28829 - Posted: 3 Aug 2009, 20:38:54 UTC - in response to Message 27734.  
Last modified: 3 Aug 2009, 20:42:39 UTC

Nice work -- by the way, do you (or anyone else here) know what the power consumption numbers comparison would be between say the 4830 and a 9800GT?

One thing I've noticed is that the 9800GT cards use a modest amount of power (though a decent 450W or 500W PS seems to work well enough with them) and quite a lot of heat.


4 setups that might help with getting to a real world nvid CUDA power draw:

C33 - Q9550@3.5GHz, Gigabyte G31 uATX mobo, Rosewill 80+ RG-430 PSU, OC'd 8800GT w/ S1 cooler - draw at the plug with full load ~210w (was 120w w/o vid card but was measured when it had an Antec EA-430 in it)

C31 - Q9550@3.4GHz, Gigabyte G31 uATX mobo, Antec 80+ EA-380 PSU, OC'd 8800GT w/ AC Turbo cooler - draw at the plug with full load ~215w (was 124w w/o vid card). Not as good of an HSF in this one. CPU and card in this one run hot. Has one extra 80mm fan in it to help blow some vid card & SB heat out the back.

C30 - Q9300@3.1GHz, Gigabyte G31 uATX mobo, Antec 80+ EA-380 PSU, OC'd 9600GSO w/ stock ASUS cooler - draw at the plug with full load ~160w (was ~106w w/o vid card), runs very cool.

C28 - Q6700@3.2GHz, Gigabyte G31 uATX mobo, Antec 80+ EA-380 PSU, 9800GT w/ stock BFG cooler - draw at the plug with full load ~209w (134w w/o vid card)

These machines have ONLY 1 small HDD, 1 large HSF, 1 small NB fan (40~50mm) and NOTHING (no cases, no case fans, no cd drives) else on them. They run in my 'server oven' that is currently 102F and often sees 107F.

Using 83% as a ballpark efficiency factor most of these things are falling real close to the sweet spot (45~50%) for PSU efficiency.

I'm gonna move the 9600GSO to a machine with a 300w, 80+ Seasonic PSU someday, time permitting as that should be a better match up and save the one with the EA-380 for some future used 8800/9800 acquisition.

EDIT: Poop! I posted this while logged onto my wife's account. Oh well.
ID: 28829 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Skip Da Shu
Avatar

Send message
Joined: 11 Apr 08
Posts: 81
Credit: 57,755,743
RAC: 0
Message 28831 - Posted: 3 Aug 2009, 20:48:05 UTC - in response to Message 27637.  
Last modified: 3 Aug 2009, 20:50:02 UTC

Guide to using the ATI MW app

A) So far, the following cards work:

4770


Did somebody get one of these working... I never did even using the 'install 8.21 and copy in 9.x files needed for this card' 4890 method. Returned it and ordered a 4850 (hmmm better check front porch, s/b here today). Really too bad because this card is arguably the price/power/output combo winner. Sure would like to hear from someone who actually has THIS card working. Then I could do some cursing about restock fees and order another one.
- da shu @ HeliOS,
"A child's exposure to technology should never be predicated on an ability to afford it."
ID: 28831 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile cornel
Avatar

Send message
Joined: 28 Feb 09
Posts: 38
Credit: 10,200,014
RAC: 0
Message 28865 - Posted: 4 Aug 2009, 11:29:44 UTC

I have an Ati HD 4770 and it's working fine on Windows 7 with Catalyst 9.7, but beware, opening CCC, Windows Media Player or GPU-Z will crash the driver. Anyway, Riva Tuner is fully compatible with this board(OC at 860/3652 Mhz).

Cornel.
ID: 28865 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
John Clark

Send message
Joined: 4 Oct 08
Posts: 1734
Credit: 64,228,409
RAC: 0
Message 28872 - Posted: 4 Aug 2009, 12:27:02 UTC - in response to Message 28831.  

Returned it and ordered a 4850 (hmmm better check front porch, s/b here today). Really too bad because this card is arguably the price/power/output combo winner. Sure would like to hear from someone who actually has THIS card working.


I run a 4850 on this rig, together with FreeHAL and Einstein. Only 3 cores are allowed to crunch Einstein to give FreeHAl and MW some core support.

I do think the 4850 should give more than the 61K to 66K RAC I have seen, but this may be due to the BOINC version (6.4.7), OS (32 bit XP Pro), etc, that I run it on/with. I think my 4850 should give me more RAC (70K to 75|K).
Go away, I was asleep


ID: 28872 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile Labbie
Avatar

Send message
Joined: 29 Aug 07
Posts: 327
Credit: 116,463,193
RAC: 0
Message 28873 - Posted: 4 Aug 2009, 12:41:47 UTC - in response to Message 28872.  

Returned it and ordered a 4850 (hmmm better check front porch, s/b here today). Really too bad because this card is arguably the price/power/output combo winner. Sure would like to hear from someone who actually has THIS card working.


I run a 4850 on this rig, together with FreeHAL and Einstein. Only 3 cores are allowed to crunch Einstein to give FreeHAl and MW some core support.

I do think the 4850 should give more than the 61K to 66K RAC I have seen, but this may be due to the BOINC version (6.4.7), OS (32 bit XP Pro), etc, that I run it on/with. I think my 4850 should give me more RAC (70K to 75|K).


My 512MB 4850 at stock get about the same as what you are on XP32, but my 1GB 4850 (overclocked) gets about 75K and my 4870 (also overclocked) gets about 80K.

You might want to check the GPU Activity load in the CCC, and if it is low, change the Wait State parameter in the cmdline to 0.8 or 0.9. My load is almost always above 85%, usually above 90%.


Calm Chaos Forum...Join Calm Chaos Now
ID: 28873 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
John Clark

Send message
Joined: 4 Oct 08
Posts: 1734
Credit: 64,228,409
RAC: 0
Message 28878 - Posted: 4 Aug 2009, 14:03:21 UTC - in response to Message 28873.  

Returned it and ordered a 4850 (hmmm better check front porch, s/b here today). Really too bad because this card is arguably the price/power/output combo winner. Sure would like to hear from someone who actually has THIS card working.


I run a 4850 on this rig, together with FreeHAL and Einstein. Only 3 cores are allowed to crunch Einstein to give FreeHAl and MW some core support.

I do think the 4850 should give more than the 61K to 66K RAC I have seen, but this may be due to the BOINC version (6.4.7), OS (32 bit XP Pro), etc, that I run it on/with. I think my 4850 should give me more RAC (70K to 75|K).


My 512MB 4850 at stock get about the same as what you are on XP32, but my 1GB 4850 (overclocked) gets about 75K and my 4870 (also overclocked) gets about 80K.

You might want to check the GPU Activity load in the CCC, and if it is low, change the Wait State parameter in the cmdline to 0.8 or 0.9. My load is almost always above 85%, usually above 90%.


I should have mentioned that my 4850 is the 512MB one running at stock.

GPU-Z says the 4850 is running at 78%, and I set the wait state to 1.25. I have now changed the value to 0.8 and GPU-Z now shows the load at 97%.

Thanks Labbie

Go away, I was asleep


ID: 28878 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Skip Da Shu
Avatar

Send message
Joined: 11 Apr 08
Posts: 81
Credit: 57,755,743
RAC: 0
Message 28880 - Posted: 4 Aug 2009, 14:21:11 UTC

Can you post the entry in context so I can see what's around it / where it goes?
ID: 28880 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile kashi

Send message
Joined: 30 Dec 07
Posts: 311
Credit: 149,490,184
RAC: 0
Message 28881 - Posted: 4 Aug 2009, 15:04:07 UTC - in response to Message 28880.  

Is this what you mean, Skip?

John Clark's HD 4850 details from Task details stderr out section:

Running Milkyway@home ATI GPU application version 0.19f by Gipsel
setting minimum kernel frequency to 20 Hz
allowing 3 concurrent WUs per GPU
scaling the wait times with 0.8

This would be achieved by inserting f20 and w0.8 in the middle of <cmdline></cmdline> in the app_info.xml file, like so:

<app_version>
<app_name>milkyway</app_name>
<version_num>19</version_num>
<flops>1.0e11</flops>
<avg_ncpus>0.1</avg_ncpus>
<max_ncpus>1</max_ncpus>
<cmdline>f20 w0.8</cmdline>
<file_ref>
ID: 28881 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Skip Da Shu
Avatar

Send message
Joined: 11 Apr 08
Posts: 81
Credit: 57,755,743
RAC: 0
Message 28883 - Posted: 4 Aug 2009, 15:47:47 UTC - in response to Message 28881.  

Is this what you mean, Skip?

John Clark's HD 4850 details from Task details stderr out section:

Running Milkyway@home ATI GPU application version 0.19f by Gipsel
setting minimum kernel frequency to 20 Hz
allowing 3 concurrent WUs per GPU
scaling the wait times with 0.8

This would be achieved by inserting f20 and w0.8 in the middle of <cmdline></cmdline> in the app_info.xml file, like so:

<app_version>
<app_name>milkyway</app_name>
<version_num>19</version_num>
<flops>1.0e11</flops>
<avg_ncpus>0.1</avg_ncpus>
<max_ncpus>1</max_ncpus>
<cmdline>f20 w0.8</cmdline>
<file_ref>


Yea, was after the app_info.xml entry as I haven't touched the one that came out of the archive file. Thanx.
ID: 28883 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile kashi

Send message
Joined: 30 Dec 07
Posts: 311
Credit: 149,490,184
RAC: 0
Message 28885 - Posted: 4 Aug 2009, 18:17:23 UTC - in response to Message 28883.  

What I posted earlier is only part of the of the app_info.xml file, not all of it. It was just to show where the <cmdline> section was located and how to configure a wait factor of 0.8. Also included the kernel frequency of 20 Hz as John Clark was using on his HD3850.

I thought it better not to post it all because the application name is different for the 64-bit version and the 32-bit version.

I now use a wait factor of 1.1 on my HD 4890 with a kernel frequency of 20 Hz. I found this reduced the temperature by 3 or 4 degrees and only increased the time to complete a task by a few seconds.
ID: 28885 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile borandi
Avatar

Send message
Joined: 21 Feb 09
Posts: 180
Credit: 27,806,824
RAC: 0
Message 28887 - Posted: 4 Aug 2009, 18:45:46 UTC - in response to Message 28872.  

Returned it and ordered a 4850 (hmmm better check front porch, s/b here today). Really too bad because this card is arguably the price/power/output combo winner. Sure would like to hear from someone who actually has THIS card working.


I run a 4850 on this rig, together with FreeHAL and Einstein. Only 3 cores are allowed to crunch Einstein to give FreeHAl and MW some core support.

I do think the 4850 should give more than the 61K to 66K RAC I have seen, but this may be due to the BOINC version (6.4.7), OS (32 bit XP Pro), etc, that I run it on/with. I think my 4850 should give me more RAC (70K to 75|K).


I believe you misunderstood his post. The 4770 is the best card in terms of price/power/output to which he was referring if anyone had got it working. As his didnt, he returned it and got a 4850.


ID: 28887 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Skip Da Shu
Avatar

Send message
Joined: 11 Apr 08
Posts: 81
Credit: 57,755,743
RAC: 0
Message 28901 - Posted: 5 Aug 2009, 2:10:49 UTC - in response to Message 28885.  

What I posted earlier is only part of the of the app_info.xml file, not all of it. It was just to show where the <cmdline> section was located and how to configure a wait factor of 0.8. Also included the kernel frequency of 20 Hz as John Clark was using on his HD3850.

I thought it better not to post it all because the application name is different for the 64-bit version and the 32-bit version.

I now use a wait factor of 1.1 on my HD 4890 with a kernel frequency of 20 Hz. I found this reduced the temperature by 3 or 4 degrees and only increased the time to complete a task by a few seconds.


I was wanting this for a 4850 I'll be installing by this weekend. Actually I have it temp installed in my wife's Windoze machine just to see it work (replacing her HD3870 for a bit). I understand the n weighting off of 1 but can you explain the f parm a bit? The HD3870 is running with app_info.xml as delivered and seems to do well with that.
ID: 28901 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile kashi

Send message
Joined: 30 Dec 07
Posts: 311
Credit: 149,490,184
RAC: 0
Message 28908 - Posted: 5 Aug 2009, 4:55:54 UTC - in response to Message 28901.  

I can't explain the technicalities of the kernel frequency any better than the readme.txt file. What it does is affects how many chunks the work unit iterations are broken up into. For example on my HD 4890 with the default kernel frequency of 30 Hz the iterations of current tasks are split into 5 chunks and with the kernel frequency set to 20 Hz, split into 4 chunks. You can see this information in your Task details stderr out section, for example: "dividing each iteration in 4 parts | borders of the domains at 0 400 800 1200 1600"

In theory less chunks should be faster to process but I have found with my HD 4890 when I had the wait state at default that it made no noticeable difference if I set f to 20 Hz, 15 Hz or 10 Hz. However when I increased the wait state to 1.1, I chose to set the kernel frequency to 20 Hz because with wait state at 1.1, kernel frequency changes make more difference for me.

My experience with the HD 4890 regarding kernel frequency settings is different than I noticed with my HD 3850. So although the principle is the same the results on different speed GPUs of experimenting with these parameters may be different. This is because the faster the GPU, the less chunks the iterations need to be split into and vice versa. The results of changing the f parameter can also be different for different task sizes. The tasks are all the same size currently but in the past there were short, long and medium length tasks and changing the f parameter from default sometimes made certain length tasks speed up and other length tasks slow down. So GPU core speed, wait state, kernel frequency and task size are all related in a complex way that sometimes means changes in different combinations of w and f parameters do not always result in the expected change in performance.

What I mean by all this, is that although you are free to try parameter settings that work for others, it is not a bad idea to try out a few different combinations yourself to see if they suit your own particular GPU and needs. Some require task processing speed above all else and do not use the computer for anything other than BOINC, some require stability, some require a more responsive desktop and some prefer to run at a reduced GPU utilisation percentage which reduces processing speed but decreases the heat load on their GPU core.

The application was well designed to run with stability and efficiency for the majority of users, so if you are happy with how it performs in an unaltered default configuration, there is no need to be concerned about changing any of these values. If you are comfortable tinkering with things that's fine too, enjoy experimenting. Retaining stability and getting the scheduling right so that BOINC debt does not build up is a more important priority in my opinion than chasing the last few seconds of speed improvement. It's little use flying through the WUs in record time if your GPU locks up every day or if MilkyWay or another project's tasks stop processing for hours at a time because of BOINC scheduling issues.
ID: 28908 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile kashi

Send message
Joined: 30 Dec 07
Posts: 311
Credit: 149,490,184
RAC: 0
Message 28911 - Posted: 5 Aug 2009, 6:42:22 UTC - in response to Message 28885.  
Last modified: 5 Aug 2009, 6:44:33 UTC

.....It was just to show where the <cmdline> section was located and how to configure a wait factor of 0.8. Also included the kernel frequency of 20 Hz as John Clark was using on his HD3850.....

Oops my apologies, not HD3850, I meant his HD 4850, the one with f20 and w0.8 parameters.
ID: 28911 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Sharon

Send message
Joined: 11 Apr 08
Posts: 10
Credit: 17,339,474
RAC: 0
Message 28956 - Posted: 6 Aug 2009, 0:26:19 UTC - in response to Message 28908.  
Last modified: 6 Aug 2009, 0:27:01 UTC

I can't explain the technicalities of the kernel frequency any better than the readme.txt file. What it does is affects how many chunks the work unit iterations are broken up into. For example on my HD 4890 with the default kernel frequency of 30 Hz the iterations of current tasks are split into 5 chunks and with the kernel frequency set to 20 Hz, split into 4 chunks. You can see this information in your Task details stderr out section, for example: "dividing each iteration in 4 parts | borders of the domains at 0 400 800 1200 1600"

In theory less chunks should be faster to process but I have found with my HD 4890 when I had the wait state at default that it made no noticeable difference if I set f to 20 Hz, 15 Hz or 10 Hz. However when I increased the wait state to 1.1, I chose to set the kernel frequency to 20 Hz because with wait state at 1.1, kernel frequency changes make more difference for me.

My experience with the HD 4890 regarding kernel frequency settings is different than I noticed with my HD 3850. So although the principle is the same the results on different speed GPUs of experimenting with these parameters may be different. This is because the faster the GPU, the less chunks the iterations need to be split into and vice versa. The results of changing the f parameter can also be different for different task sizes. The tasks are all the same size currently but in the past there were short, long and medium length tasks and changing the f parameter from default sometimes made certain length tasks speed up and other length tasks slow down. So GPU core speed, wait state, kernel frequency and task size are all related in a complex way that sometimes means changes in different combinations of w and f parameters do not always result in the expected change in performance.

What I mean by all this, is that although you are free to try parameter settings that work for others, it is not a bad idea to try out a few different combinations yourself to see if they suit your own particular GPU and needs. Some require task processing speed above all else and do not use the computer for anything other than BOINC, some require stability, some require a more responsive desktop and some prefer to run at a reduced GPU utilisation percentage which reduces processing speed but decreases the heat load on their GPU core.

The application was well designed to run with stability and efficiency for the majority of users, so if you are happy with how it performs in an unaltered default configuration, there is no need to be concerned about changing any of these values. If you are comfortable tinkering with things that's fine too, enjoy experimenting. Retaining stability and getting the scheduling right so that BOINC debt does not build up is a more important priority in my opinion than chasing the last few seconds of speed improvement. It's little use flying through the WUs in record time if your GPU locks up every day or if MilkyWay or another project's tasks stop processing for hours at a time because of BOINC scheduling issues.


Most Kewlness! EXACTLY the level and sort of 'explaination' I was after.

I swapped out the HD3870 in my wifes desktop, put in my new HD4850 and saw the typically ~3:30 WUs drop t around 1:59 each. But GPU-Z also showed the GPU was only seeing a load ranging between 67% and 77%. That seemed a bit low to me and I wanted to see if there was a way to work it a little harder because it's eventual target machine is a dedicated number cruncher. That's why I was asking and that's what I am playing with.

It appears I have the load range now in the 83 to 97% range using no "f" parm but using "w0.9". I also clocked GPU up from it's default of 625 to 675Mhz. The WUs times have dropped to something closer to 1:29.

I had tried "w0.8" & a clock rate of 690MHz and that cut the WU times by another 6~8 seconds but I also saw the GPU core temp running 3~5C higher. I set it back to "w0.9" and the clock back to 665 for now. Seems to be 'happy' settings for it in general.

I'll play around with the "f" parm a bit once it gets into it's target machine (waiting on a huge Aqua to finish so I can re-OS that machine).

You are right about the scheduler. Last night a couple times after a restart I had to suspend all the other projects for 1 second to get MW to actually start running or to start running 3 WUs even though the resource share is set such that it's share is 73% and the next largest share is 5%. Haven't had that occur today. Looks like it's back into it's happy routine of running 3, 1 waiting to report, 2 in que. In a couple minutes it'll finish 2, start the two in que, report 1~3, get 2 or 3 more and start the cycle over again.

Dang I sure wish MW and Cullzart would port these CAL apps over to Linux 64b.

Anyway, thanx much for your time.
ID: 28956 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Skip Da Shu
Avatar

Send message
Joined: 11 Apr 08
Posts: 81
Credit: 57,755,743
RAC: 0
Message 28957 - Posted: 6 Aug 2009, 0:28:47 UTC - in response to Message 28956.  

Dang, I posted using her account again...
ID: 28957 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile verstapp
Avatar

Send message
Joined: 26 Jan 09
Posts: 589
Credit: 497,834,261
RAC: 0
Message 28965 - Posted: 6 Aug 2009, 5:09:15 UTC

We'll just call you Sharon from now on. :)
Cheers,

PeterV

.
ID: 28965 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Skip Da Shu
Avatar

Send message
Joined: 11 Apr 08
Posts: 81
Credit: 57,755,743
RAC: 0
Message 28993 - Posted: 7 Aug 2009, 8:53:26 UTC
Last modified: 7 Aug 2009, 8:54:24 UTC

Has anyone tried running BOINC w/ the MW ATI app in a WinXP VirtualBox under Debian/Ubuntu?

signed, Sharon LOL
ID: 28993 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Skip Da Shu
Avatar

Send message
Joined: 11 Apr 08
Posts: 81
Credit: 57,755,743
RAC: 0
Message 29005 - Posted: 7 Aug 2009, 16:16:28 UTC - in response to Message 28993.  

Has anyone tried running BOINC w/ the MW ATI app in a WinXP VirtualBox under Debian/Ubuntu?

signed, Sharon LOL


How about WinXP 32b in VirtualBox under any 64b Linux?
ID: 29005 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Previous · 1 · 2 · 3 · 4 · 5 · Next

Message boards : Number crunching : Using the MW ATi application, Q & A

©2024 Astroinformatics Group