WUs completing too quickly & getting too few WUs
log in

Advanced search

Message boards : Number crunching : WUs completing too quickly & getting too few WUs

1 · 2 · Next
Author Message
David A. Dutton [TopGun]
Send message
Joined: 16 Oct 11
Posts: 16
Credit: 109,325,124
RAC: 0
Message 21686 - Posted: 9 Nov 2015, 21:40:59 UTC
Last modified: 9 Nov 2015, 21:42:46 UTC

I only use the GPU setting, and prefs have CUDA and OPENCL checked. When I uncheck Open GL and just leave CUDA, I get absolutely NO WUs. IT seems they are open CL only... but that's no biggie.. I just leave both checked, and I also check the 2 apps for Collatz as well...

I can set BOINC to 10 days + 10 days MORE worth of work, and I can only get about 60 or so WUs... I manually edited CC config and told boinc that the max WUs to DL for any project was 500... but most projects seem to have a 50-100 WU limit...

My issue is that my PC works on 4 WUs at once, and each WU finishes in 5mins... so in 60 minutes, I complete 48 WUs Per hour. That means everything I DL finishes in about 1.5 hrs or less.My other machines take longer - especially the 580 cards and the 8800M laptop.. that takes almost an hr to do 1 WU. But the Titan Cards let me change many of the GPU processors over to double FP to speed up Cuda and open cl programs

Of course I Can leave the project setting in boinc to always get new tasks so they don't run out, but I typically like to DL WUs before bed, and get enough so they finish in about 8 hrs, and then im on my PC most of the day and boinc is off.

Is 60 or so the HARD limit of WUs one can have for this Project?
____________
Box 1:
* Intel i7-2600K OCed to 5GHz, H20
* Win 10 X64 - 16GB DDR3 RAM @ 1666Mhz
* 2, NV Titan Z cards in Quad SLI mode
* C:1.5TB,D:7TB,F:24TB

Box 2: GTX 680 X2 SLI
Box 3: GTX 580X2 SLI
Box 4: GTX 780TIx2 SLI
Box 5: Older Laptop 8800Mx2 SLI

Profile Slicker
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist
Avatar
Send message
Joined: 11 Jun 09
Posts: 2525
Credit: 740,580,099
RAC: 1
Message 21687 - Posted: 10 Nov 2015, 2:36:05 UTC

Every time a host contacts the project server, it provides the list of WUs it has so the server can compare to the list it has and see if there are any missing. If so, it resends them. The alternative is to let the WU timeout. But, when people tweak the various xml files (cc_config, app_config, etc) and somethine goes wrong, it is possible that the client will delete the WUs without marking them as having errored out. Checking all those WUs every time, especially when people have their machines set to contact the server as often as possible, causes a huge load on the server. So, the better choice would be to have the same or even fewer WUs but increase the size. While you will likely take the time to adjust your preferences to run only the large WUs (if they existed), other's would complain that they get large WUs when they didn't choose them and now they've crunched one for a week or two and it will never finish in time. In other words, you can't please all of the people all of the time. That being said, I think it is time to add a large sieve app for fast GPUs which means setting up a validator, work generator, and assimilator.

numbermaniac
Send message
Joined: 26 Jul 14
Posts: 30
Credit: 6,014,522
RAC: 13,700
Message 21688 - Posted: 10 Nov 2015, 5:32:11 UTC - in response to Message 21687.

Every time a host contacts the project server, it provides the list of WUs it has so the server can compare to the list it has and see if there are any missing. If so, it resends them. The alternative is to let the WU timeout. But, when people tweak the various xml files (cc_config, app_config, etc) and somethine goes wrong, it is possible that the client will delete the WUs without marking them as having errored out. Checking all those WUs every time, especially when people have their machines set to contact the server as often as possible, causes a huge load on the server. So, the better choice would be to have the same or even fewer WUs but increase the size. While you will likely take the time to adjust your preferences to run only the large WUs (if they existed), other's would complain that they get large WUs when they didn't choose them and now they've crunched one for a week or two and it will never finish in time. In other words, you can't please all of the people all of the time. That being said, I think it is time to add a large sieve app for fast GPUs which means setting up a validator, work generator, and assimilator.


If that's what it used to be before the Sieve, that should work now too. Different work unit sizes would be optimal, because one size doesn't fit all here.

numbermaniac
Send message
Joined: 26 Jul 14
Posts: 30
Credit: 6,014,522
RAC: 13,700
Message 21689 - Posted: 10 Nov 2015, 5:33:52 UTC - in response to Message 21686.

I only use the GPU setting, and prefs have CUDA and OPENCL checked. When I uncheck Open GL and just leave CUDA, I get absolutely NO WUs. IT seems they are open CL only... but that's no biggie.. I just leave both checked, and I also check the 2 apps for Collatz as well...

I can set BOINC to 10 days + 10 days MORE worth of work, and I can only get about 60 or so WUs... I manually edited CC config and told boinc that the max WUs to DL for any project was 500... but most projects seem to have a 50-100 WU limit...

My issue is that my PC works on 4 WUs at once, and each WU finishes in 5mins... so in 60 minutes, I complete 48 WUs Per hour. That means everything I DL finishes in about 1.5 hrs or less.My other machines take longer - especially the 580 cards and the 8800M laptop.. that takes almost an hr to do 1 WU. But the Titan Cards let me change many of the GPU processors over to double FP to speed up Cuda and open cl programs

Of course I Can leave the project setting in boinc to always get new tasks so they don't run out, but I typically like to DL WUs before bed, and get enough so they finish in about 8 hrs, and then im on my PC most of the day and boinc is off.

Is 60 or so the HARD limit of WUs one can have for this Project?

A few minutes? WTF is with your computers ?! My tasks take ~60 minutes on the Nvidia GPU, ~2-3 hours on Intel GPU and up to 30 hours on CPU. How do all of you get those tasks done in minutes?!!!!!

David A. Dutton [TopGun]
Send message
Joined: 16 Oct 11
Posts: 16
Credit: 109,325,124
RAC: 0
Message 21690 - Posted: 10 Nov 2015, 8:19:45 UTC - in response to Message 21687.

Thats my problem collatz doesn't have long running gpu apps.. or if it does im not getting them.... or I might be.. 2 Titan Zs is 4 GPus with over 20+K processors... plus I have them set to mostly double float .. I have to turn that off for gaming or gaming goes to crap...

The server only sends me ONE type of WU (although there's an option for 2)

I get Collatz Sieve 1.21 opencl nvidia gpu.. (I guess that's the MINI version)... about a year ago I was getting LONGER WUs for my GPUs that would run for a lot longer...
____________
Box 1:
* Intel i7-2600K OCed to 5GHz, H20
* Win 10 X64 - 16GB DDR3 RAM @ 1666Mhz
* 2, NV Titan Z cards in Quad SLI mode
* C:1.5TB,D:7TB,F:24TB

Box 2: GTX 680 X2 SLI
Box 3: GTX 580X2 SLI
Box 4: GTX 780TIx2 SLI
Box 5: Older Laptop 8800Mx2 SLI

David A. Dutton [TopGun]
Send message
Joined: 16 Oct 11
Posts: 16
Credit: 109,325,124
RAC: 0
Message 21691 - Posted: 10 Nov 2015, 8:51:58 UTC - in response to Message 21689.

I don't use CPU for Collatz just the 4 GPUs on my 2 Titan Zs. The Titan Zs are GREAT cards with 6GB per GPU (12 per card)... BUT they are clocked ridiculously slow...

in a nutshell a Titan Z is pretty much 2 GTX 780TIs on one card..

But remember I think im only getting the MINI collatz WUs.. sieve 1.21, open_cl_nvidia_gpu

BUT the titan Z cards have a driver setting to turn most of the processing units inside each GPU into dedicated double precision FPUs.... most NV cards couldn't do this back then.. I imagine most can today...I don't own any 800/900 line of cards... no point yet... with 2 T Zs in quad SLI.. I can run ANY game at 2560x1600 at max qualities without a blip

plus I OC the hell out of them... the T Zs have a TERRIBLY low clock speed.. but they are basically 2 780TIs on one card (im over simplifying it a bit)....

With a VERY CREATIVE H20 cooling system, my CPU runs at 4.9-5Ghz which is why I LOVE my older intel i7-2600K and I have all the Tz overclocked by 500Mhz and set Precision EVGA to allow the cards to draw up to 120% as needed... I can also put them in a mode where it can go beyond 120% and tell it temp is the priority...


here is how i cool.. I yanked out the h20 cooling system in this alienware r3 unit... built a custom one...

the problem is most H20 units have ONE pump that circles water to all devices and that does always get cool water to devices down the line... AND that water must be COOLED BY air in a "radiator".. the problem is that Alienware machines use the air from INSIDE the box to cool the water while sucking air out.. my main cooling fan is in the FRONT... air draws in from the floor (where it's the coolest)... then the fan blasts all the PC cards so THEIR fans can pulling cooler air....even then the heat of these was keeping my case at 90F in a 66F home office (i keep it 65f year round)...

then I got an idea... I bought a portable 16,000 BTU AC and it has a window kit for the exhaust pipe.. one system pulls in room air and passes over the cooling "radiator" (condenser/evaporator.. I always get confused which is which named)... then it has other intakes that pull in room and that cool another "radiator" to blow the heat off them and return the coolant nice and cold while pushing the air out a tube , and out by window...


So... I make my own H20 system.. with one , strong pump... that splits off into 5 plastic tubes... CPU and 2 to to parts on the cooling plate for each card... then all 5 recombine into 1 tube which i then run out of the PC (I 100% disabled the CPU cooling fan/radiator)... that tube attaches to a connector that splits the flow into 20 SMALL copper pipes which I inserted into the 16K BTU and physically attached them to the "radiator" that gets COLD (about 40F).. plus the after the air passes through it, it's around 45F.... these small, METAL tubes bleed off the heat FAST.. sure it reduces the room cooling of the unit a bit.. but only by 2K BTU or so.. and that's MORE than enough.. 16K portable unit could cool an entire house floor of 1000sq ft...

anyway the water then comes back to a metal connector that is inserted into a plastic tube.. it recombines all the water from the tine copper tubes.. the pump then splits it into 5 more tubes so EACH device gets COLD water that NO other device warms.... each device is hit with 42-45F... Hence why I OC things so high...

At 5GHZ with the CPU under massive load, its temp is 65C... each GPU ides at 10-15C... and when under massive load, each GPU never goes over 70C.. and these cards are MADE to run long term at 80-85C... TTROTTLE is still set to never let the CPU get over 73C and the GPUs 82C.... this also has REALLY reduced the ambient temp inside my box from 90F (when everything was running hot)... to a cool 75F when under load... in place of the old CPU radiator, I put in a fan and connected it to the port that regulated the radiator fan to cool the CPUs water as the system wont boot unless it detects a running fan..

but I paid about nearly $4000 for all these GPUs....

If anyone doesn't believe how fast i crank out the WUs I an screenshot boincs ETA to finish them.. but it never gets it right.. it said 00:06:01 per WU.. but it really takes 4:50 to 5:01

and for CPU tasks I use milkyway at home because it sends you multicore WUs based on the number of ACTIVE CPUs you have sent in boinc.. so I get (6CPUs) next to each WU for MW.. I rarely use all CPUs because it's really 4 CPUS that can handle 2 threads and that's NOT quite the same as 4 more REAL CPUs.. but I need cPU time thats free for even GPU apps as they need SMALL cpu time...

If I overload my CPU, those 5M WUs can take up to 30mins... The CPU is need to feed data to the GPU and or system ram....if it's bogged down.. this doesn't happen.. plus I'm also a meologiet and have a weather system on my room and I pull data ever 2 seconds.. but it only need 0.5-1% total CPU use at all times


As far as someone mentioning errors, I don't get any.. unless I do something to
____________
Box 1:
* Intel i7-2600K OCed to 5GHz, H20
* Win 10 X64 - 16GB DDR3 RAM @ 1666Mhz
* 2, NV Titan Z cards in Quad SLI mode
* C:1.5TB,D:7TB,F:24TB

Box 2: GTX 680 X2 SLI
Box 3: GTX 580X2 SLI
Box 4: GTX 780TIx2 SLI
Box 5: Older Laptop 8800Mx2 SLI

Profile mikey
Avatar
Send message
Joined: 11 Aug 09
Posts: 3246
Credit: 1,700,925,421
RAC: 4,845,256
Message 21692 - Posted: 10 Nov 2015, 11:27:46 UTC - in response to Message 21689.


WTF is with your computers ?! My tasks take ~60 minutes on the Nvidia GPU, ~2-3 hours on Intel GPU and up to 30 hours on CPU. How do all of you get those tasks done in minutes?!!!!!


The easy answer is that your GTX 740 only has CUDA Cores: 384, while his Titan has CUDA Cores: 5760, the more Cuda cores the faster your gpu crunches, add in that he has 12gb of memory while yours has has between 2 and 4gb of on board ram and his gpu is just more powerful than yours is.

My own AMD 7970 which has 3gb of on board ram and 2048 Cuda cores estimates it will take 3:46 to finish one unit. My other gpu here is my Nvidia 560Ti which has 384 Cuda cores and 1gb of on board memory is doing the units in just under 15 minutes each.

In short while your Nvidia 740 is plenty capable the newer gpu's are faster and more powerful. To make yours run faster are you leaving a cpu core free just to feed the gpu? If not try it and see if that helps, sometimes it does and sometimes the effect is minimal, each system is different.

Old man
Send message
Joined: 24 Aug 09
Posts: 782
Credit: 14,832,359
RAC: 0
Message 21695 - Posted: 10 Nov 2015, 14:23:05 UTC - in response to Message 21692.
Last modified: 10 Nov 2015, 14:23:51 UTC

My own AMD 7970 which has 3gb of on board ram and 2048 Cuda cores


You have wonderful AMD gpu then. :) Or write error.

woohoo
Send message
Joined: 2 Oct 15
Posts: 8
Credit: 1,739,300,411
RAC: 1,325,961
Message 21696 - Posted: 10 Nov 2015, 14:26:30 UTC

Where do they sell those at?

Profile Crunch3r
Volunteer moderator
Project developer
Project tester
Avatar
Send message
Joined: 30 Jun 09
Posts: 219
Credit: 7,515,663,568
RAC: 10,847
Message 21697 - Posted: 10 Nov 2015, 17:15:12 UTC - in response to Message 21696.

Where do they sell those at?


It's Mikey.. so my guess would be "la la land".
____________

Team BOINC United.Join Science that matters.

Anthony Ayiomamitis
Send message
Joined: 21 Jan 15
Posts: 48
Credit: 1,095,123,072
RAC: 10,078,946
Message 21698 - Posted: 10 Nov 2015, 20:35:23 UTC - in response to Message 21692.


WTF is with your computers ?! My tasks take ~60 minutes on the Nvidia GPU, ~2-3 hours on Intel GPU and up to 30 hours on CPU. How do all of you get those tasks done in minutes?!!!!!

Pretty much my case as well with my dinosaur Intel HD 4000 which requires about 75 minutes on average per work unit (+ 0.544 cores). I am not really complaining - in fact, I love it - since this the secondary GPU in my MSI GE70 and which nets me around 80k credits per day.

I also have an NVidia GeForce GTX 660M (yes, slightly dated) but I cannot get it to work with Collatz after we moved on from V6.04 with work units bombing after the first second or two. Oh how nice it would be to also have my 660M as well running Collatz and similar to when we were at V6.04!

Profile mikey
Avatar
Send message
Joined: 11 Aug 09
Posts: 3246
Credit: 1,700,925,421
RAC: 4,845,256
Message 21701 - Posted: 11 Nov 2015, 11:49:21 UTC - in response to Message 21695.

My own AMD 7970 which has 3gb of on board ram and 2048 Cuda cores


You have wonderful AMD gpu then. :) Or write error.


Okay, okay :stream processors", does that make you feel better, Nvidia calls them cuda cores, AMD calls them stream processors. They are NOT comparable between the brands but they ARE comparable among the same brand.

Profile mikey
Avatar
Send message
Joined: 11 Aug 09
Posts: 3246
Credit: 1,700,925,421
RAC: 4,845,256
Message 21702 - Posted: 11 Nov 2015, 11:51:20 UTC - in response to Message 21696.

Where do they sell those at?


Amazon...http://www.amazon.com/Sapphire-Radeon-PCI-Express-Graphics-11197-01-40G/dp/B007JGARVU/ref=sr_1_3?ie=UTF8&qid=1447242617&sr=8-3&keywords=amd+7970+gpu

and my guess is many other places too

Profile sosiris
Send message
Joined: 11 Dec 13
Posts: 123
Credit: 55,800,869
RAC: 0
Message 21709 - Posted: 12 Nov 2015, 7:52:14 UTC - in response to Message 21691.

Don't worry. I'm currently working on a more computing intensive (at least 2.5x) sieve app which adopts the 256-step sieve instead of the 32-step one currently used. And it generates 2.5 million times more 'science' than the one without sieve. BTW, it also consumes much less GPU RAM which is a good news.
____________
Sosiris, team BOINC@Taiwan

Anthony Ayiomamitis
Send message
Joined: 21 Jan 15
Posts: 48
Credit: 1,095,123,072
RAC: 10,078,946
Message 21713 - Posted: 12 Nov 2015, 12:02:43 UTC - in response to Message 21702.

Where do they sell those at?


Amazon...http://www.amazon.com/Sapphire-Radeon-PCI-Express-Graphics-11197-01-40G/dp/B007JGARVU/ref=sr_1_3?ie=UTF8&qid=1447242617&sr=8-3&keywords=amd+7970+gpu

and my guess is many other places too

Mike, which BOINC projects benefit particularly well from this card? I know it eats up Collatz work units especially well.

Helix Von Smelix
Send message
Joined: 2 Aug 10
Posts: 43
Credit: 10,195,934,378
RAC: 1,484,771
Message 21722 - Posted: 13 Nov 2015, 19:30:44 UTC

400+ Seconds for a WU on the Titan Z. I would send them back.
I am getting 60-70 seconds on GTX980Ti and R9 200 AMD Hawaii

Not sure why your Titan Z is so slow?

numbermaniac
Send message
Joined: 26 Jul 14
Posts: 30
Credit: 6,014,522
RAC: 13,700
Message 21723 - Posted: 13 Nov 2015, 20:49:09 UTC - in response to Message 21722.

400+ Seconds for a WU on the Titan Z. I would send them back.
I am getting 60-70 seconds on GTX980Ti and R9 200 AMD Hawaii

Not sure why your Titan Z is so slow?


Well, he did say it was severely underclocked...

numbermaniac
Send message
Joined: 26 Jul 14
Posts: 30
Credit: 6,014,522
RAC: 13,700
Message 21724 - Posted: 13 Nov 2015, 20:51:59 UTC - in response to Message 21692.


WTF is with your computers ?! My tasks take ~60 minutes on the Nvidia GPU, ~2-3 hours on Intel GPU and up to 30 hours on CPU. How do all of you get those tasks done in minutes?!!!!!


The easy answer is that your GTX 740 only has CUDA Cores: 384, while his Titan has CUDA Cores: 5760, the more Cuda cores the faster your gpu crunches, add in that he has 12gb of memory while yours has has between 2 and 4gb of on board ram and his gpu is just more powerful than yours is.

My own AMD 7970 which has 3gb of on board ram and 2048 Cuda cores estimates it will take 3:46 to finish one unit. My other gpu here is my Nvidia 560Ti which has 384 Cuda cores and 1gb of on board memory is doing the units in just under 15 minutes each.

In short while your Nvidia 740 is plenty capable the newer gpu's are faster and more powerful. To make yours run faster are you leaving a cpu core free just to feed the gpu? If not try it and see if that helps, sometimes it does and sometimes the effect is minimal, each system is different.


I only use 2/4 cores for BOINC, so there should be still be one or two for the GPU.

numbermaniac
Send message
Joined: 26 Jul 14
Posts: 30
Credit: 6,014,522
RAC: 13,700
Message 21725 - Posted: 13 Nov 2015, 20:53:00 UTC - in response to Message 21709.

Don't worry. I'm currently working on a more computing intensive (at least 2.5x) sieve app which adopts the 256-step sieve instead of the 32-step one currently used. And it generates 2.5 million times more 'science' than the one without sieve. BTW, it also consumes much less GPU RAM which is a good news.


Sounds promising!

Profile BarryAZ
Send message
Joined: 21 Aug 09
Posts: 251
Credit: 13,330,398,603
RAC: 23,638,499
Message 21726 - Posted: 14 Nov 2015, 4:53:31 UTC - in response to Message 21692.

What I've been running on the Nvidia side are 750, and 750ti.

The 750ti cards push out Collatz work units a bit slower than my ATI 7850.

1 · 2 · Next
Post to thread

Message boards : Number crunching : WUs completing too quickly & getting too few WUs


Main page · Your account · Message boards


Copyright © 2018 Jon Sonntag; All rights reserved.