All sky GW search taking "All Day" to "All Week"

JStateson
JStateson
Joined: 7 May 07
Posts: 67
Credit: 510,011,424
RAC: 864,796
Topic 219603

Sorry, but I had to abort them.  Will exclude them in the future using "venue".  There is also huge inconsistency in credits but the main problem (for me) is inconsistency in expected time to finish.  From a day to a week on an S9000 coprocessor that normally takes 20 minutes for the GPB search.

 

Eskomorko
Eskomorko
Joined: 15 Jan 09
Posts: 30
Credit: 373,026,313
RAC: 164,693

Same here. My RAC drops like

Same thing happened here. My RAC drops like tenfold every time when my computer crunches these.

Jim1348
Jim1348
Joined: 19 Jan 06
Posts: 342
Credit: 176,423,674
RAC: 58,901

I am not quite sure what an

I am not quite sure what an S9000 is, but the All-Sky v1.07 are remarkably consistent on my RX 570 (Win7 64-bit).   They are all within a minute or two of 74 minutes.

Richie
Richie
Joined: 7 Mar 14
Posts: 431
Credit: 1,523,386,424
RAC: 375,157

Jim1348 wrote:I am not quite

Jim1348 wrote:
I am not quite sure what an S9000 is

"Its price at launch was 2499 US Dollars"
https://www.techpowerup.com/gpu-specs/firepro-s9000.c1879

Jim1348
Jim1348
Joined: 19 Jan 06
Posts: 342
Credit: 176,423,674
RAC: 58,901

Richie wrote:"Its price at

Richie wrote:
"Its price at launch was 2499 US Dollars"

Thanks. I probably wouldn't buy one now.

Eskomorko
Eskomorko
Joined: 15 Jan 09
Posts: 30
Credit: 373,026,313
RAC: 164,693

Jim1348 wrote:I am not quite

Jim1348 wrote:
I am not quite sure what an S9000 is, but the All-Sky v1.07 are remarkably consistent on my RX 570 (Win7 64-bit).   They are all within a minute or two of 74 minutes.

With 2070RTX it usually takes 90-110mins per wu and CPU ones approx. 12-14hrs per wu. RAC was about 300 000 before and now it's 50 000 and still going downhill fast. Running 24/7 but sometimes it looks like no progress is made.

JStateson
JStateson
Joined: 7 May 07
Posts: 67
Credit: 510,011,424
RAC: 864,796

Jim1348 wrote:Richie

Jim1348 wrote:
Richie wrote:
"Its price at launch was 2499 US Dollars"

Thanks. I probably wouldn't buy one now.

 

$84 on ebay free ship (roll your own cooler)

Double precision float speed is 806 GLOPs

rx570 is 318

gtx1070ti is only 255

 

It will do 5 concurrent Milkyway work units in 3.5 minutes.  Unfortunately, when it runs out of data 900 WU's later, the project takes 10-15 minutes before it sends new data so my fallback task is Einstein.  

Milkyway makes extensive use of double precision GPU hardware.  Not sure why, probably expedient.   Originally I thought the app was coded in Fortran, but no, they used C.

Jim1348
Jim1348
Joined: 19 Jan 06
Posts: 342
Credit: 176,423,674
RAC: 58,901

Then if I had a S9000, I

Then if I had a S9000, I would use it on MilkyWay, and just let it idle when they don't send work.  I don't see how it can be so slow on Einstein.  There must be something wrong.  Maybe it needs more CPU support?  (My ASRock RX 570 was $150. at Newegg.)

The RX 570 does MilkyWay in about 103 seconds for a single WU; I could probably do 2 and get that down to maybe 80 or 90 on average.  That is about the limit of what the MW servers will provide.

 

JStateson
JStateson
Joined: 7 May 07
Posts: 67
Credit: 510,011,424
RAC: 864,796

Jim1348 wrote: There must be

Jim1348 wrote:
There must be something wrong.  Maybe it needs more CPU support?

I suspect the problem is the x1 -> x16 risers in this mining rig.  The i7-4790s and 8gb ram should be OK.  The CPU is also running Einstein's OpenCL-intel-gpu-beta and takes only 11 minutes for a "BRP".  The ATI OpenCL "GPB" on the s9000 takes min 20 minutes.   I am not sure if that comparison is valid as the apps are different, but if it is then the risers add a minimum of 9 minutes.  I assume those "all sky" must not be suitable for running on risers due to restricted bandwidth of the x1 but that is a guess.  The "full day" to a "full week" might be calculated that way due to Einstein running at lower priority than the milkyway.  clearly I should not b e running those tasks as they wont finish so I aborted them to allow them to be resent to other users.

I just ran a test on three gtx1070 TI using one GPB per gpu and then two concurrent GPBs.  Things were worse which surprised me.

12.57 00:12:34 D2
12.83 00:12:50 D1
12.60 00:12:36 D0
12.42 00:12:25 D2
29.52 00:29:31 D1
29.42 00:29:25 D2
29.53 00:29:32 D1
30.63 00:30:38 D0
30.68 00:30:41 D0
29.48 00:29:29 D2
12.45 00:12:27 D2
12.58 00:12:35 D0
12.67 00:12:40 D1
12.60 00:12:36 D0
12.48 00:12:29 D2

This desktop system did not have any risers and running a pair of concurrent work units took an extra 5 minutes over running the pair in tandem. GPUz showed %95 load with just one work unit so there was no room for an additional task.  I stopped all other tasks when running the test.

Jim1348
Jim1348
Joined: 19 Jan 06
Posts: 342
Credit: 176,423,674
RAC: 58,901

JI suspect the problem is the

Quote:
I suspect the problem is the x1 -> x16 risers in this mining rig.  The i7-4790s and 8gb ram should be OK.

That could be it.  I am impressed with the DP performance, and see why you like it.  If MW ever got their act together, they would fix their servers to take advantage of cards like that.  But that is one reason it is best for me to stay here.  They don't need me on MW, they have plenty of very good cards to work with.

DanNeely
DanNeely
Joined: 4 Sep 05
Posts: 1,279
Credit: 1,349,478,450
RAC: 1,222,080

I have no idea what's behind

I have no idea what's behind your inconsistent runtimes; but even when working properly the current GW GPU app produces much lower credit numbers than the Fermi GPU app.  Hopefully this situation will improve as new versions of the app are developed; in the mean time you can opt out of it by going into your project preferences and setting 'Run test applications?:" to No.

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.