Join PrimeGrid
Returning Participants
Community
Leader Boards
Results
Other
drummers-lowrise
|
Message boards :
Project Staging Area :
wwww: Percent of GPU time waiting for GPU: xx.xx
| Author |
Message |
|---|
|
|
|
Hi,
I'm playing with wwww.ini trying to optimize wwww (wwwwcl64.exe with an OpenCL Intel gfx)
How to tune thread/blocksize? What the meaning of Percent of GPU time waiting for GPU: xx.xx?
Should xx.xx tend to 100% or 0%?
For wieferich + opencl, it's better to have more thread? bigger block?
Thanks.
____________
Badge Score: 1*2 + 1*5 + 12*6 + 4*7 + 2*8 + 1*9 = 132 | |
|
rogueVolunteer developer
Send message
Joined: 8 Sep 07
Posts: 1249
ID: 12001
Credit: 18,565,548
RAC: 0
 
|
|
As blocksize increases you will see more lag in updates to your display.
For Wieferich a larger percentage of time is spent sieving than for Wall-Sun-Sun.
When trying to determine the optimal setting, I suggest that you make sure that nothing else is running then run wwwwcl with various values for the number of threads and blocksize and focus on the ones that complete the fastest (overall run time). You can then adjust the numbers further to see if you can do better. | |
|
|
|
|
I perform some tests. I create a bat and run the following command
wwwwcl64.exe -p 10000e10 -P 10001e10 -T Wieferich -t [1-5] -b [1000-5000].
Results below. There is not an obvious winner. On current wieferich range, my p/sec is around 3.6M
t b p/sec time
1 1000 4.407M 70.66 seconds
2 1000 4.377M 71.41 seconds
3 1000 4.392M 70.69 seconds
4 1000 4.420M 70.42 seconds
5 1000 4.398M 70.72 seconds
1 2000 4.411M 70.66 seconds
2 2000 4.403M 70.70 seconds
3 2000 4.428M 70.42 seconds
4 2000 4.294M 72.05 seconds
5 2000 4.440M 70.33 seconds
1 3000 4.415M 70.58 seconds
2 3000 4.439M 70.42 seconds
3 3000 4.382M 71.05 seconds 4 3000 4.416M 70.31 seconds 5 3000 4.424M 70.45 seconds
1 4000 4.396M 71.17 seconds
2 4000 4.312M 71.88 seconds
3 4000 4.355M 71.33 seconds
4 4000 4.367M 71.49 seconds
5 4000 4.216M 73.22 seconds
1 5000 4.327M 71.94 seconds
2 5000 4.416M 70.35 seconds
3 5000 4.369M 71.32 seconds
4 5000 4.374M 71.08 seconds
5 5000 4.350M 71.70 seconds
Do you think about anything else I could try to find optimal config?
Thanks
____________
Badge Score: 1*2 + 1*5 + 12*6 + 4*7 + 2*8 + 1*9 = 132 | |
|
Scott Brown Volunteer moderator
Project administrator
Volunteer tester
Project scientist
Send message
Joined: 17 Oct 05
Posts: 2329
ID: 1178
Credit: 15,637,975,109
RAC: 10,119,420
                                           
|
I perform some tests. I create a bat and run the following command
wwwwcl64.exe -p 10000e10 -P 10001e10 -T Wieferich -t [1-5] -b [1000-5000].
Results below. There is not an obvious winner. On current wieferich range, my p/sec is around 3.6M
t b p/sec time
1 1000 4.407M 70.66 seconds
2 1000 4.377M 71.41 seconds
3 1000 4.392M 70.69 seconds
4 1000 4.420M 70.42 seconds
5 1000 4.398M 70.72 seconds
1 2000 4.411M 70.66 seconds
2 2000 4.403M 70.70 seconds
3 2000 4.428M 70.42 seconds
4 2000 4.294M 72.05 seconds
5 2000 4.440M 70.33 seconds
1 3000 4.415M 70.58 seconds
2 3000 4.439M 70.42 seconds
3 3000 4.382M 71.05 seconds 4 3000 4.416M 70.31 seconds 5 3000 4.424M 70.45 seconds
1 4000 4.396M 71.17 seconds
2 4000 4.312M 71.88 seconds
3 4000 4.355M 71.33 seconds
4 4000 4.367M 71.49 seconds
5 4000 4.216M 73.22 seconds
1 5000 4.327M 71.94 seconds
2 5000 4.416M 70.35 seconds
3 5000 4.369M 71.32 seconds
4 5000 4.374M 71.08 seconds
5 5000 4.350M 71.70 seconds
Do you think about anything else I could try to find optimal config?
Thanks
Your p/sec is about where I would expect it for an onboard GPU (for example, a discrete GT640 card that should solidly outperform any Intel gfx runs about 10M p/sec). I am, however, surprised by your times that seem far too fast for that p/sec rate. For example, an OEM GTX660 card can run around 32M p/sec, but that equates to only about 85 sec per unit. The GT640 is more like 265 sec per unit. At your p/sec rate, I would expect something more along the lines of 600 to 700 seconds. Are your measuring that from the "clock time" output or something else?
| |
|
|
|
|
Scott,
He tests with -p 10000e10 -P 10001e10 which is a shorter interval than one of the WU from PRPNet port 13000.
/JeppeSN | |
|
Scott BrownVolunteer moderator
Project administrator
Volunteer tester
Project scientist
Send message
Joined: 17 Oct 05
Posts: 2329
ID: 1178
Credit: 15,637,975,109
RAC: 10,119,420
|
Scott,
He tests with -p 10000e10 -P 10001e10 which is a shorter interval than one of the WU from PRPNet port 13000.
/JeppeSN
Doh!...note to self to stop reading and replying when I am in a rush posted to forehead. :)
| |
|
|
|
|
I ran the test with 'small range'.
I should have specified, the graphic card is an on board Intel HD 4600.
When running WFS on port 1300, playing with the t&b settings, the time per WU is between 691 and 703 sec. (around 3.4-3.6 M p/sec)
So, if there is no such differences, is there a 'better' setting? Higher t? Higher b? No matter?
PS: Also, as in Title, what means 'Percent of GPU time waiting for GPU'? Lower is better? Higher is better?
Should it be; Percent of CPU time waiting for GPU?
Thanks.
____________
Badge Score: 1*2 + 1*5 + 12*6 + 4*7 + 2*8 + 1*9 = 132 | |
|
Message boards :
Project Staging Area :
wwww: Percent of GPU time waiting for GPU: xx.xx |
