Join PrimeGrid
Returning Participants
Community
Leader Boards
Results
Other
drummers-lowrise
|
Message boards :
Project Staging Area :
WWWCL + Intel iGPU?
| Author |
Message |
|---|
|
|
|
So, I've been experimenting with WWWCL and Intel iGPU. In specific, the HD 4000 which is OpenCL and DirectCompute 5.0 capable.
I' ve noticed that starting wwwcl task does not actually use the GPU until CTRL+C is invoked in the command window. I can see this with GPUz.
Before CTRL+C is invoked, GPUz shows 0-1% GPU utilization. After CTRL+C is invoked, the command windows shows "CTRL-C accepted. Please wait for threads to completed" at which point GPUz starts to show 30-50% GPU utilization.
Is this a bug in the wwwcl program? Is it an known issue?
Also, CTRL+C doesn't always gracefully shutdown wwwcl, I had 1 task become abandoned due to the program not completing the WU before shutting down.
Thoughts?
____________
1281*2^594565+1
2393323632147*2^1290000-1 | |
|
|
|
|
Something interesting of note. The WU I started earlier today finished, looking in the test_results.log file I see that the checksum is "0000000000000000" is this normal? Or is it an erroneous result?
Also, are wwwwcl WUs supposed to start at 50% ?
My prpclient.log shows:
[2016-07-06 12:04:31 PDT] WIEFERICH: Getting work from server prpnet.primegrid.com at port 13000
[2016-07-06 12:04:33 PDT] WIEFERICH: PRPNet server is version 5.4.0
[2016-07-06 13:50:03 PDT] : Could not parse stats line in [wwww.log]
[2016-07-06 13:51:43 PDT] PRPNet Client application v5.4.0 started
[2016-07-06 13:51:43 PDT] User name Ion_Basa at email address is ******(omitted)
[2016-07-06 15:26:23 PDT] WIEFERICH: Range 511656000000000000 to 511656100000000000 completed
[2016-07-06 15:26:23 PDT] Total Time: 1:34:41 Total Work Units: 1 Special Results Found: 0
[2016-07-06 15:26:24 PDT] WIEFERICH: Returning work to server prpnet.primegrid.com at port 13000
[2016-07-06 15:26:24 PDT] WIEFERICH: INFO: Test for range 511656000000000000:511656100000000000 was accepted
[2016-07-06 15:26:24 PDT] WIEFERICH: INFO: All 1 test results were accepted
I don't want to be returning bad results to the server. Can someone please clarify?
____________
1281*2^594565+1
2393323632147*2^1290000-1 | |
|
 Sysadm@Nbg Volunteer moderator
Volunteer tester
Project scientist
Send message
Joined: 5 Feb 08
Posts: 1178
ID: 18646
Credit: 449,776,341
RAC: 125,589
                    
|
|
seems okay to me
done the test offline:
[07.07.2016 17:20:57][sysadm@ubuntu-i7-adv:~/prpclient-5.4.0a-linux_64/prpclient-2]
$ ./wwwwcl -v -p 511656000000000000 -P 511656100000000000 -T Wieferich -b 3072 -t 2
wwwwcl v2.2.5, a GPU program to search for Wieferich and WallSunSun primes
Compiled Jun 12 2013 with GCC 4.6.1
Platform 0 is a NVIDIA Corporation NVIDIA CUDA, version OpenCL 1.2 CUDA 8.0.20
Device 0 is a NVIDIA Corporation GeForce GTX 970
workGroupSize = 1277952 = 3072 * 32 * 13 (blocks * workGroupSizeMultiple * deviceComputeUnits)
Running with 2 threads
Allocated memory (prior to sieving): 29 MB in CPU, 29 MB in GPU
Sieve started: (cmdline) 511656000000000000 <= p < 511656100000000000
p=511656071549093957, 29.20M p/sec, 1.69 CPU cores, 71.5% done. ETA 07 Jul 17:22
Sieve complete: 511656000000000001 <= p < 511656100000000000 2452349706 primes tested
Clock time: 84.77 seconds at at 28930802 p/sec.
Processor time: 142.67 sec. (0.71 init + 141.96 sieve).
Seconds spent in CPU and GPU: 3.08 (cpu), 60.44 (gpu)
Percent of time spent in CPU vs. GPU: 4.85 (cpu), 95.15 (gpu)
CPU/GPU utilization: 1.68 (cores), 0.71 (devices)
[07.07.2016 17:22:51][sysadm@ubuntu-i7-adv:~/prpclient-5.4.0a-linux_64/prpclient-2]
$ cat wwww.log
Primes tested 2452349706. Checksum 0000000000000000. Time 84.766047 seconds
____________
Sysadm@Nbg
my current lucky number: 3749*2^1555697+1
PSA-PRPNet-Stats-URL: http://u-g-f.de/PRPNet/
 | |
|
|
|
|
Good to know. I was worried because the CPU program returns a non-zero checksum.
So it seems that Intel iGPU graphics does work (now). There was talk back in 2013 that the HD graphics in Ivy Bridge was lacking double precision support due to it being lacking in the drivers. I'm not sure if WWWWCL requires DP or if Intel pushed out a driver that enabled OpenCL DP at some point.
Something else to note:
Earlier I said GPU utilization doesn't happen until CTRL+C is invoked, this turns out to be wrong. It just takes awhile for the WU to start on the GPU. Usually by the time the WU reaches 1-2%. It appears for Wall-Sun-Sun after the initial 1-2% it jumps to 50%. Is this normal?
____________
1281*2^594565+1
2393323632147*2^1290000-1 | |
|
|
|
seems okay to me
I tried with the (slow) CPU program:
.\wwww.exe -p511656000000000000 -P511656100000000000 -tWieferich -s1000 -v
wwww v1.3, a program to search for Wieferich and Wall-Sun-Sun primes.
Searching for Wieferich primes between 511656000000000000 and 511656100000000000.
The largest prime in the sieve is 715301999 (using 37003736 primes) using
approximately 182677 KB of memory.
Searched to 511656100010175120 (100.00 done)
The search completed in 8363.82 seconds (306501.70 tests per second)
Spent 2598.69 seconds sieving and 5675.77 seconds testing
Tested 2563524092 values (2.56 pct) of the range
Primes tested 2563524092. Checksum 0000000014e6c82a. Time 8363.817000
The last line above is also an entry in wwww.log afterwards.
Is it to be expected that the CPU app has another checksum? I think an all-zero checksum (as in the GPU app) seems improbable, unless there is an explanation.
/JeppeSN | |
|
|
|
seems okay to me
I tried with the (slow) CPU program:
.\wwww.exe -p511656000000000000 -P511656100000000000 -tWieferich -s1000 -v
wwww v1.3, a program to search for Wieferich and Wall-Sun-Sun primes.
Searching for Wieferich primes between 511656000000000000 and 511656100000000000.
The largest prime in the sieve is 715301999 (using 37003736 primes) using
approximately 182677 KB of memory.
Searched to 511656100010175120 (100.00 done)
The search completed in 8363.82 seconds (306501.70 tests per second)
Spent 2598.69 seconds sieving and 5675.77 seconds testing
Tested 2563524092 values (2.56 pct) of the range
Primes tested 2563524092. Checksum 0000000014e6c82a. Time 8363.817000
The last line above is also an entry in wwww.log afterwards.
Is it to be expected that the CPU app has another checksum? I think an all-zero checksum (as in the GPU app) seems improbable, unless there is an explanation.
/JeppeSN
Sysadm@Nbg Tested it with the wwwwcl app on a gtx 970 and he too had an all zero checksum. You're asking the same question I'm trying to ask too, why the all zero checksum on the GPU app, but not the CPU app? I'm not quite sure why that would be the case, which is why I am asking here!
____________
1281*2^594565+1
2393323632147*2^1290000-1 | |
|
|
|
Sysadm@Nbg Tested it with the wwwwcl app on a gtx 970 and he too had an all zero checksum. You're asking the same question I'm trying to ask too, why the all zero checksum on the GPU app, but not the CPU app? I'm not quite sure why that would be the case, which is why I am asking here!
You are right, you already said yourself that the CPU checksum was different. So I agree my post is just a repetition of the question you had.
And I think it is a good question. Why do the checksums not agree, and why is one of them (the GPU one) zero?
/JeppeSN | |
|
rogueVolunteer developer
Send message
Joined: 8 Sep 07
Posts: 1191
ID: 12001
Credit: 18,565,548
RAC: 0
 
|
Sysadm@Nbg Tested it with the wwwwcl app on a gtx 970 and he too had an all zero checksum. You're asking the same question I'm trying to ask too, why the all zero checksum on the GPU app, but not the CPU app? I'm not quite sure why that would be the case, which is why I am asking here!
You are right, you already said yourself that the CPU checksum was different. So I agree my post is just a repetition of the question you had.
And I think it is a good question. Why do the checksums not agree, and why is one of them (the GPU one) zero?
Because I never completed the code to properly create a checksum that could be used for double-checking. | |
|
|
|
Because I never completed the code to properly create a checksum that could be used for double-checking.
Ahh, so as it stands there's no way of us double checking GPU tasks with one another. Since you coded wwwwcl , did you ever test on Intel iGPUs? Is there any way to run a manual test on my own client for testing purposes?
____________
1281*2^594565+1
2393323632147*2^1290000-1 | |
|
rogueVolunteer developer
Send message
Joined: 8 Sep 07
Posts: 1191
ID: 12001
Credit: 18,565,548
RAC: 0
|
Because I never completed the code to properly create a checksum that could be used for double-checking.
Ahh, so as it stands there's no way of us double checking GPU tasks with one another. Since you coded wwwwcl , did you ever test on Intel iGPUs? Is there any way to run a manual test on my own client for testing purposes?
Yes. Run the wwwwcl program from the command line with -h (IIRC) to list the arguments it accepts. Use -t (threads) and -b (blocks) to find optimal throughput. These can also be get in wwwwcl.ini when the program is run as part of PRPNet. | |
|
|
|
|
Of course you can run on an interval where there is a known near-hit and verify that you re-find it. /JeppeSN | |
|
Message boards :
Project Staging Area :
WWWCL + Intel iGPU? |
