Freeing CPU cores

Message boards : Number crunching : Freeing CPU cores
Message board moderation

To post messages, you must log in.

Previous · 1 · 2 · 3 · 4 · Next

AuthorMessage
Cruncher-American Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor

Send message
Joined: 25 Mar 02
Posts: 1513
Credit: 370,893,186
RAC: 340
United States
Message 1401200 - Posted: 9 Aug 2013, 19:53:54 UTC

In general, here's a quick way to check if you need to free up a core - i.e, if your CPU is overcommitted:

1) Check ratio of CPU time to elapsed time (for CPU WUs and AP on GPU). If the CPU time is typically < 90%, free up a core.

2) (if in Windows): look at Performance tab in Task Manager. Turn on Show Kernel Times under View. If the red is consistently more than 10-20%, your cores are fighting with each other because there are too many threads running (remember, BOINC isn't the only thing running on your computer - Windows needs some room, too). Free up a core.
ID: 1401200 · Report as offensive
Profile John Neale
Volunteer tester
Avatar

Send message
Joined: 16 Mar 00
Posts: 634
Credit: 7,246,513
RAC: 9
South Africa
Message 1401246 - Posted: 9 Aug 2013, 22:26:00 UTC - in response to Message 1401092.  

<snip>

Your laptops have a NVS 4200M not sure but i belive it´s a pre-Fermi GPU, so you can´t run more than 1 WU at a time. It´s hard to say why you get only 10% of GPU usage, my bet is in some power saving setting feature normaly used by the laptop makers to save batery and/or avoid overheat. Look at the energy usage on the control pannel that maybe give you a clue. Try to change from economy to high performance and see if it help. But remember, not push to hard a laptop, heat is their enemy, and GPU crunch produces a lot of heat.

An NVS 4200M is definitely not pre-Fermi. I've been running two MB v6 and v7 work units using the optimised apps on my laptop's NVS 4200M GPU for a long time, with no adverse effects. More recently, I've also started running two AP at a time. It does get a bit warm, more so (a) in summer, (b) when running AP and (c) when plugged into my stupidly-designed docking station that blocks half of the air vent, but I use TThrottle to keep the CPU and GPU temperatures under 75 °C.
ID: 1401246 · Report as offensive
Profile Mike Special Project $75 donor
Volunteer tester
Avatar

Send message
Joined: 17 Feb 01
Posts: 34249
Credit: 79,922,639
RAC: 80
Germany
Message 1401249 - Posted: 9 Aug 2013, 22:53:53 UTC

Very interesting John.
Thank you for this information.
Very useful.



With each crime and every kindness we birth our future.
ID: 1401249 · Report as offensive
juan BFP Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 16 Mar 07
Posts: 9786
Credit: 572,710,851
RAC: 3,799
Panama
Message 1401256 - Posted: 9 Aug 2013, 23:36:57 UTC - in response to Message 1401246.  

<snip>

Your laptops have a NVS 4200M not sure but i belive it´s a pre-Fermi GPU, so you can´t run more than 1 WU at a time. It´s hard to say why you get only 10% of GPU usage, my bet is in some power saving setting feature normaly used by the laptop makers to save batery and/or avoid overheat. Look at the energy usage on the control pannel that maybe give you a clue. Try to change from economy to high performance and see if it help. But remember, not push to hard a laptop, heat is their enemy, and GPU crunch produces a lot of heat.

An NVS 4200M is definitely not pre-Fermi. I've been running two MB v6 and v7 work units using the optimised apps on my laptop's NVS 4200M GPU for a long time, with no adverse effects. More recently, I've also started running two AP at a time. It does get a bit warm, more so (a) in summer, (b) when running AP and (c) when plugged into my stupidly-designed docking station that blocks half of the air vent, but I use TThrottle to keep the CPU and GPU temperatures under 75 °C.

That´s good, so you could give GS a hand and show how to configurate the laptop to rise it´s gpu usage. Now is on your hands.

ID: 1401256 · Report as offensive
Profile Raistmer
Volunteer developer
Volunteer tester
Avatar

Send message
Joined: 16 Jun 01
Posts: 6325
Credit: 106,370,077
RAC: 121
Russia
Message 1401447 - Posted: 10 Aug 2013, 12:41:17 UTC - in response to Message 1401115.  


That may be just luck due to the nature of the work units, or it may be due to some change in the driver with 13.8. I will let it continue to run to see if it is reliable.


It's quite interesting. What blanking % those tasks have?

SETI apps news
We're not gonna fight them. We're gonna transcend them.
ID: 1401447 · Report as offensive
Profile cov_route
Avatar

Send message
Joined: 13 Sep 12
Posts: 342
Credit: 10,270,618
RAC: 0
Canada
Message 1401462 - Posted: 10 Aug 2013, 14:27:34 UTC - in response to Message 1401447.  
Last modified: 10 Aug 2013, 14:31:05 UTC


That may be just luck due to the nature of the work units, or it may be due to some change in the driver with 13.8. I will let it continue to run to see if it is reliable.


It's quite interesting. What blanking % those tasks have?

It varies widely, 0 to 80%. The high-blanking ones take maybe 70% longer to run* which doesn't surprise me.

Edit: *Except for a couple of 100% blanked which terminated after a few seconds.
ID: 1401462 · Report as offensive
Profile Raistmer
Volunteer developer
Volunteer tester
Avatar

Send message
Joined: 16 Jun 01
Posts: 6325
Credit: 106,370,077
RAC: 121
Russia
Message 1401464 - Posted: 10 Aug 2013, 14:31:50 UTC - in response to Message 1401462.  

Well, pay attention to possible correlation between increase in memory page faults (as you saw on graphs in other thread) and increase in number of "misses" in app's counters (they exist both for AP and MB). Does such correlation exist or not?...
SETI apps news
We're not gonna fight them. We're gonna transcend them.
ID: 1401464 · Report as offensive
Profile cov_route
Avatar

Send message
Joined: 13 Sep 12
Posts: 342
Credit: 10,270,618
RAC: 0
Canada
Message 1401585 - Posted: 10 Aug 2013, 21:00:17 UTC - in response to Message 1401464.  

Well, pay attention to possible correlation between increase in memory page faults (as you saw on graphs in other thread) and increase in number of "misses" in app's counters (they exist both for AP and MB). Does such correlation exist or not?...

I am waiting for the system slowdown to happen again, I have some instrumentation running to catch it. It's been a few days since it happened.
ID: 1401585 · Report as offensive
Profile James Sotherden
Avatar

Send message
Joined: 16 May 99
Posts: 10436
Credit: 110,373,059
RAC: 54
United States
Message 1404164 - Posted: 17 Aug 2013, 6:01:42 UTC

Tuesday I freed up one core and ran computer ID 6814791 with 7 cores with one feeding the GPU. I ran that way untill 1:00 pm friday afternoon.

The result was I lost a tad over 1,000 in rac. compared to my control computer ID 7003180. These two were within 200 Rac at the start of the test.

These computers are identical in all componets but memory. Comp ID 6814791 has 8GB and the other has 16GB.

So for MY computer 6814791 running HT on 7 cores and running 1 WU per GPU I lost RAC by going to 7 cores.

I realize that others mileage may vary. So this weekend I need to clean the dust bunnys out. So when I power back up I will go into BIOS and turn HT off and see what happens on comp ID 6814791. I am sure I will lose more rac. But we will see.
[/quote]

Old James
ID: 1404164 · Report as offensive
Profile Mike Special Project $75 donor
Volunteer tester
Avatar

Send message
Joined: 17 Feb 01
Posts: 34249
Credit: 79,922,639
RAC: 80
Germany
Message 1404199 - Posted: 17 Aug 2013, 7:08:16 UTC

You have got very fast CPU`s there James.
You would probably benefit only by running multiple instances on your GPU.
A friend of mine is running 3 instances on a 450 and RAC increased by 1500 on his rather old CPU.
Dont forget tho.
Those tests should be made in long terms.
Lets say 4 weeks.




With each crime and every kindness we birth our future.
ID: 1404199 · Report as offensive
Profile James Sotherden
Avatar

Send message
Joined: 16 May 99
Posts: 10436
Credit: 110,373,059
RAC: 54
United States
Message 1404218 - Posted: 17 Aug 2013, 7:52:37 UTC - in response to Message 1404199.  

You have got very fast CPU`s there James.
You would probably benefit only by running multiple instances on your GPU.
A friend of mine is running 3 instances on a 450 and RAC increased by 1500 on his rather old CPU.
Dont forget tho.
Those tests should be made in long terms.
Lets say 4 weeks.



That is a good point. I will run 7 cores later on and see what happens.

I only run one on the gpu beacause temps are good and its stable.

I will run without HT for a week though. Just to see what happens.


[/quote]

Old James
ID: 1404218 · Report as offensive
Profile James Sotherden
Avatar

Send message
Joined: 16 May 99
Posts: 10436
Credit: 110,373,059
RAC: 54
United States
Message 1404520 - Posted: 18 Aug 2013, 2:25:32 UTC

Ok the dust bunnys are gone out of both my control and test machines.
On power up I went into BIOS and disabled HT for test computer 6814791. Current rac is 15,550.
Its an I7 3770 with 8 GB of ram. And a 550Ti gpu running 1 WU. no overclocking or free core to gpu.

The control computer is 7003180. Another I7 3770 but with 16 GB ram, running HT on and no free core to the 550Ti Gpu, also running 1 WU with no overclocking.
Curerent Rac is 15,891.

So lets see what happens. I will run for a week.So some time next Saturday I will end the test and see if the old saw of running a real 4 cores is better than running 8 with HT.

I predict my rac will end up in the toilet. That old saw might have been true in the old P4 days of HT. But lets find out.
[/quote]

Old James
ID: 1404520 · Report as offensive
Profile Wiggo
Avatar

Send message
Joined: 24 Jan 00
Posts: 34744
Credit: 261,360,520
RAC: 489
Australia
Message 1404603 - Posted: 18 Aug 2013, 8:15:31 UTC
Last modified: 18 Aug 2013, 8:21:15 UTC

After much testing I have come to this conclusion with my nvidia assisted rigs rigs;

Only doing cuda MB's, my 2500K can easily handling either 3x 9800Gt's (single task each) or 2x GTX550Ti's being fed 2 workunits each without needing a core free.

My Q6600 didn't either running the same video cards, but putting 2x GTX660's in it being fed 3 workunits each did require me to reserve a CPU core to feed those cards (feed times for each GPU workunit increased by 20sec's without a reserved core and core times increased greatly).

Both the 2500K & 3570K lose production output freeing a core while just doing cuda workunits.

After replacing the Q6600 with a 3570K, I'm back to using all 4 cores for the best performance (there is some small impact to CPU task times, but no where near enough to effect my everyday use or the maximum output from the rig).

It is nice though to have a second video de/en coding, editing rig again (the old Q6600 was definitely showing its age there too).

Cheers.
ID: 1404603 · Report as offensive
Profile James Sotherden
Avatar

Send message
Joined: 16 May 99
Posts: 10436
Credit: 110,373,059
RAC: 54
United States
Message 1404924 - Posted: 19 Aug 2013, 4:30:20 UTC
Last modified: 19 Aug 2013, 4:30:37 UTC

After one day of running with ht off on Computer 6814791. Starting RAC was 15,550 At this moment it is 15,541, I thought it would really drop off faster.

Control rig computer 7003180 rac at the start was 15891.It is 15753 now.

Test will continue.
[/quote]

Old James
ID: 1404924 · Report as offensive
Profile Raistmer
Volunteer developer
Volunteer tester
Avatar

Send message
Joined: 16 Jun 01
Posts: 6325
Credit: 106,370,077
RAC: 121
Russia
Message 1404960 - Posted: 19 Aug 2013, 8:58:12 UTC

RAC is too unreliable indicator of performance. One should gather statistics for task completion times vs AR (MB) and vs blanking % (for AP) to define host performance. RAC depends on few non-performance factors that makes it inappropriate tool. Using RAC you could get some performance estimates only. If expected performance differencies would be quite big then it will be reflected in RAC. In other case observed differencies in RAC well in error boundaries of such method.

SETI apps news
We're not gonna fight them. We're gonna transcend them.
ID: 1404960 · Report as offensive
Profile James Sotherden
Avatar

Send message
Joined: 16 May 99
Posts: 10436
Credit: 110,373,059
RAC: 54
United States
Message 1407516 - Posted: 25 Aug 2013, 5:22:46 UTC - in response to Message 1404960.  

RAC is too unreliable indicator of performance. One should gather statistics for task completion times vs AR (MB) and vs blanking % (for AP) to define host performance. RAC depends on few non-performance factors that makes it inappropriate tool. Using RAC you could get some performance estimates only. If expected performance differencies would be quite big then it will be reflected in RAC. In other case observed differencies in RAC well in error boundaries of such method.

That is most likely true. But as Im not good at math or being a BOINC geek I will go by RAC.
I was going to test for a week but I will go for a month.

Control computer 7003180 RAC for this day after one week is 16,303
test computer 6814791 is 15,404

So 3 weeks to go.
[/quote]

Old James
ID: 1407516 · Report as offensive
Profile Wiggo
Avatar

Send message
Joined: 24 Jan 00
Posts: 34744
Credit: 261,360,520
RAC: 489
Australia
Message 1407574 - Posted: 25 Aug 2013, 10:09:55 UTC

I've found that also monitoring the rig's total tasks in progress against your RAC can cut your test time down to around 2 weeks.

If both RAC and total tasks are still dropping after 2 weeks you then know the test has failed.

For your next test why not try running 2 MB tasks on that 550 and see what happens. Under SETI V6 this gave me a 50-60% increase in production on my 550's, but I'm not prepared to test the difference myself under SETI V7 as I know that there will still be an improvement, how much I can't say now (but I'll guess that at least a 25-30% gain should be guaranteed). ;-)

But while running an opencl AP on a 550Ti, I'm not sure that I'd run 1 MB at the same time (but then I only do AP's CPU's).

Cheers.
ID: 1407574 · Report as offensive
Profile William Kendrick

Send message
Joined: 25 Dec 08
Posts: 46
Credit: 180,614
RAC: 0
United States
Message 1407620 - Posted: 25 Aug 2013, 15:04:46 UTC
Last modified: 25 Aug 2013, 15:18:20 UTC

i didn't read it here, but i've got 6 cores, and to free an entire one i must use 82-86%. just figured id add that here since i hadn't seen it
there's still plenty of ppl using hexacores you know :P
this would actually be more usefull if you could find a way to do this for our spus on nvidia gtx gpu.

especially for those of us with nvidia cards that use amd cpus

ID: 1407620 · Report as offensive
Profile William Kendrick

Send message
Joined: 25 Dec 08
Posts: 46
Credit: 180,614
RAC: 0
United States
Message 1407621 - Posted: 25 Aug 2013, 15:07:36 UTC

also me and hal had a huge discussion about this: rac is meaningless to computer computational ability. period.. 100%. if you want a better indicator, get a cpu benchmark program, or, while not as reliable as a full blown benchmark you can test your cpu's with the boinc client. its under advanced "run cpu benchmarks" for your gpu's you can use any of the products that test those.. futurmark makes a particularly good one. :)

ID: 1407621 · Report as offensive
Profile William Kendrick

Send message
Joined: 25 Dec 08
Posts: 46
Credit: 180,614
RAC: 0
United States
Message 1407623 - Posted: 25 Aug 2013, 15:16:02 UTC - in response to Message 1401246.  

<snip>

I use TThrottle to keep the CPU and GPU temperatures under 75 °C.


whoa. i tried to look up your gpu, but i couldn't find a temperature threshold anyone agreed upon. 75C is far far too hot. i'm running an old gtx 260 core 216 and i can barely after a full night get it to 119F, how in gods name is that thing still functioning? thats an insane amount of heat to put near any components imho.

ID: 1407623 · Report as offensive
Previous · 1 · 2 · 3 · 4 · Next

Message boards : Number crunching : Freeing CPU cores


 
©2024 University of California
 
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.