My worthless non-GPU adventure


log in

Advanced search

Message boards : Number crunching : My worthless non-GPU adventure

Previous · 1 · 2 · 3 · 4 · 5 · Next
Author Message
Profile Ex
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 12 Mar 12
Posts: 2895
Credit: 1,797,699
RAC: 396
United States
Message 1239469 - Posted: 1 Jun 2012, 15:19:59 UTC - in response to Message 1239415.

Boinc has its own settings to control cpu use and heat !
Just use BOINC`s own `preferences` setting, something like :- `on multi processors use at most 50% cpu`
And use optimized app`s
It works well here.

I did something similar on my i7 before I upgraded from the stock heatsink. Except I use the cc_config option to specify the number of CPUs to 4. Later I found running 4 tasks on that machine it produced a RAC very near my other i7 boxes, with the same i7-860 CPU, that run 8 tasks at once.

As far as non-GPU machines being worthless I am currently running 1 GPU. An old GT 8500. With my position by RAC being ~80 and position by credit ~130 I wouldn't say non-GPU machines are worthless.


Hell any of us could compete with GPUs if we had as many CPU's running at our disposal as you do. ;-)
____________
-Dave #2

3.2.0-33

Profile HAL9000
Volunteer tester
Avatar
Send message
Joined: 11 Sep 99
Posts: 4600
Credit: 121,607,199
RAC: 44,878
United States
Message 1239479 - Posted: 1 Jun 2012, 15:39:12 UTC - in response to Message 1239469.

Boinc has its own settings to control cpu use and heat !
Just use BOINC`s own `preferences` setting, something like :- `on multi processors use at most 50% cpu`
And use optimized app`s
It works well here.

I did something similar on my i7 before I upgraded from the stock heatsink. Except I use the cc_config option to specify the number of CPUs to 4. Later I found running 4 tasks on that machine it produced a RAC very near my other i7 boxes, with the same i7-860 CPU, that run 8 tasks at once.

As far as non-GPU machines being worthless I am currently running 1 GPU. An old GT 8500. With my position by RAC being ~80 and position by credit ~130 I wouldn't say non-GPU machines are worthless.


Hell any of us could compete with GPUs if we had as many CPU's running at our disposal as you do. ;-)

Most of them are pretty old kit though. Nearly 75% of my RAC comes from just 6 of those machines. I could power on more but the older machines to tend to make one heck of a racket. Plus with the rack of servers we moved in here I'm not sure our A/C could handle the extra heat.

A newer $500-600 machine would probably replace 2 or 3 of my faster ones. I might get lucky and be able to justify a few new machines to "support Windows 8", but that is really up to the bean counters.
____________
SETI@home classic workunits: 93,865 CPU time: 863,447 hours

Join the BP6/VP6 User Group today!

Profile Ex
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 12 Mar 12
Posts: 2895
Credit: 1,797,699
RAC: 396
United States
Message 1239770 - Posted: 2 Jun 2012, 1:03:52 UTC

Someone familiar with the optimized apps over at arkayn? I think I picked right, but I wanna be sure...

Machine 1)
Running Linux 64 bit (on this machine anyways)
my Xeon has the SSSE3 instruction set (note THREE S's). So I found the only app there supporting SSSE3x64 (LX64_AKV8_AP6_AVX_SSSE3.zip)

I don't run a GPU, so no CUDA.

Am I all set to go with this package?
.....................................................
Machine 2)
Running Linux 32 bit. Processor only has SSE2.
I cannot decide between:
ap_6.01r546_sse2_linux32.zip
LX32_AKV8_AP6_SSE2.zip
I think ap_6.01* is just an AP app, and LX32_AK* is a package for MB and AP, am I correct? If that's the case then want the LX32 package I believe.

so for machine 2 I want LX32_AK**?


Are my thoughts here in good order before I go dumping these apps on my computers? (I cannot wait to see the kinda performance I get with these apps. The SSSE3 should be pretty sweet?)



____________
-Dave #2

3.2.0-33

Profile Ex
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 12 Mar 12
Posts: 2895
Credit: 1,797,699
RAC: 396
United States
Message 1239938 - Posted: 2 Jun 2012, 4:13:19 UTC

LOL, I'm goin' for it. I'll see if there's any issue.
____________
-Dave #2

3.2.0-33

Josef W. SegurProject donor
Volunteer developer
Volunteer tester
Send message
Joined: 30 Oct 99
Posts: 4334
Credit: 1,113,795
RAC: 779
United States
Message 1239941 - Posted: 2 Jun 2012, 4:17:28 UTC - in response to Message 1239770.
Last modified: 2 Jun 2012, 4:32:28 UTC

Someone familiar with the optimized apps over at arkayn? I think I picked right, but I wanna be sure...

Machine 1)
Running Linux 64 bit (on this machine anyways)
my Xeon has the SSSE3 instruction set (note THREE S's). So I found the only app there supporting SSSE3x64 (LX64_AKV8_AP6_AVX_SSSE3.zip)

I don't run a GPU, so no CUDA.

Am I all set to go with this package?
.....................................................
Machine 2)
Running Linux 32 bit. Processor only has SSE2.
I cannot decide between:
ap_6.01r546_sse2_linux32.zip
LX32_AKV8_AP6_SSE2.zip
I think ap_6.01* is just an AP app, and LX32_AK* is a package for MB and AP, am I correct? If that's the case then want the LX32 package I believe.

so for machine 2 I want LX32_AK**?


Are my thoughts here in good order before I go dumping these apps on my computers? (I cannot wait to see the kinda performance I get with these apps. The SSSE3 should be pretty sweet?)

Arkayn's out of town for a few days, so although I'm no expert on the Linux apps I'm offering this advice. If someone actually using Linux disagrees, I suggest you take their advice.

Yes, for machine 2 the LX32_AKV8_AP6_SSE2.zip package is correct.

Editting out the red stuff....
For machine 1, the LX64_AKV8_AP6_AVX_SSSE3.zip contains an AVX application for AP_v6 which probably won't work. The LX64_AKV8_AP6_SSE3.zip package would be safer.

The ideal combination would be the SSE3 application for AP_v6 and the SSSE3 application for MB. That would take some editing of app_info.xml, etc. If you want to go that way we can work out the details, I'm too close to bedtime to be sure I'd get it right just now.


I forgot to look at your hosts before posting that. Since the XEON is a Sandy Bridge which does have AVX, you're probably good to go with the package containing the AVX AP_v6 app. You're also running a very recent Linux so there's no issue with the OS having the needed support.
Joe

Profile Ex
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 12 Mar 12
Posts: 2895
Credit: 1,797,699
RAC: 396
United States
Message 1239977 - Posted: 2 Jun 2012, 5:00:32 UTC
Last modified: 2 Jun 2012, 5:01:49 UTC

Both boxes took the apps and started right back up. I fear the Xeon bunked a few WU's in the transfer, (my list got wayyyy shorter)...


But both boxes appear to be in shape and running. I'll be watching for valid results.


And Mark, I will probably do my windows laptop as well while I'm running it. Can you tell me how to find out what instructions my processor supports, in windows????

(That's an easy one in linux, but damned if I know in windows.....)


And Thanks Josef.... You see something about my processor that linux didn't tell me apparently... :-) (hence some of my confusion)
____________
-Dave #2

3.2.0-33

Profile Ex
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 12 Mar 12
Posts: 2895
Credit: 1,797,699
RAC: 396
United States
Message 1239991 - Posted: 2 Jun 2012, 5:17:25 UTC

typical windoze situation. Download this and use it... Lol


In linux, all you do is type: "less /proc/cpuinfo"

[ok, done making a political case]



Thanks Mark, I will do just that, I appreciate the solution. :-D
____________
-Dave #2

3.2.0-33

Profile Raistmer
Volunteer developer
Volunteer tester
Avatar
Send message
Joined: 16 Jun 01
Posts: 3588
Credit: 48,759,314
RAC: 21,535
Russia
Message 1239997 - Posted: 2 Jun 2012, 5:25:18 UTC - in response to Message 1239249.

Just wondering why you think optimised applications are of no use. They will work regardless of whether or not a CPU is throttled and they should be more efficient per clock/time-slice/whatever-measure-of-time-you-want-to-use. I'm not saying you have to use the optimised applications, but your reason for not doing so doesn't seem to make sense for me.


I've heard from others that optimized apps don't do much for throttled systems. If there would still be a noticeable improvement then perhaps I'll start taking steps to do so, but that means more learning (and I do enough of that all day on my own).

Another concern I have is heat. Do optimized apps produce more heat? Or perhaps less because they are "optimized"? (Then again it really doesn't matter because I'll just throttle my temps to taste. :-))

So I guess it's time to learn about optimized apps for Linux... (Probably won't bother with my windows laptop, it'll only be running boinc for a few weeks)


At least AP optimization includes high-level optimizations that will benefit at any throttled CPUs cause opt app just does less computations that stock one. Of course, another part that opt app does remaining amount more quickly hence bigger CPU temp. But one just can throttle CPU more if temp under considerations.

Profile Ex
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 12 Mar 12
Posts: 2895
Credit: 1,797,699
RAC: 396
United States
Message 1240035 - Posted: 2 Jun 2012, 7:15:21 UTC
Last modified: 2 Jun 2012, 7:22:53 UTC

Wow, first WU's in from my Xeon under optimized apps, looks like 2x the credit per processing time...

Just WOW.

I'll be watching this for a couple days to see what really happens.


This just begs the question, why are optimized apps not standard? Just too much trouble making sure the right apps go to the right architectures?


And just for a benchmark: I started this with RAC under 2000 (Due to the outage and having no workunits processing then) My normal RAC running just the xeon at 30-40%, was between 2000 and 2500.

We will see how high it goes. I think I'll hit that goal of 7,000 with the added machines and op apps.
____________
-Dave #2

3.2.0-33

Profile Ex
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 12 Mar 12
Posts: 2895
Credit: 1,797,699
RAC: 396
United States
Message 1240061 - Posted: 2 Jun 2012, 8:38:34 UTC

After watching several units finish under op apps on my good xeon box, I can safely report that I am getting roughly 2.5X the credit per proc cycle....

Again, all I can say is WOW. I'm very pleased. I may break 10,000 on this adventure.
____________
-Dave #2

3.2.0-33

Profile Ex
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 12 Mar 12
Posts: 2895
Credit: 1,797,699
RAC: 396
United States
Message 1240467 - Posted: 3 Jun 2012, 0:27:43 UTC
Last modified: 3 Jun 2012, 0:32:15 UTC

So my current RAC of 2,600 is already an all time personal high. Just 24 hours into this or so... :-) Cool, I'll just be watching this grow.



But what is MOST impressive to me, is my Xeon linux machine which I switched from stock to optimized apps. The difference in work done/processor time, is INSANE :-D

For comparison I found two units with almost identical process time-
Stock app:
proc time credit
5,386.45 34.66

Optimized app: (SSSE3)
proc time credit
5,388.34 115.28

Basically, I tripled my credit potential on the Xeon box!, not to mention I'm currently giving it twice as much CPU time as I was, so I expect "nano" to end up doing 12,000 all by itself (when my norm was 2000)

Talk about wow, I'm still amazed at the difference with the op app. I should thank whoever designed it!, and thanks to Arkayn's site for having it right there for me!
____________
-Dave #2

3.2.0-33

Profile perryjay
Volunteer tester
Avatar
Send message
Joined: 20 Aug 02
Posts: 3377
Credit: 16,211,348
RAC: 5,406
United States
Message 1240687 - Posted: 3 Jun 2012, 16:50:03 UTC - in response to Message 1240467.

Ex, I wouldn't look for that kind of credit change per work unit. There are too many variables such as the amount of blanking, angle range, credit new (which is a pain), and also your wingman and the amount of credit he claims. The main thing to look for with the opt-apps is the completion times compared to what you were doing before on work with the same angle ranges. It's a pretty good bet you will find you are getting at least half again or close to twice the speed you were before.
____________


PROUD MEMBER OF Team Starfire World BOINC

Profile Ex
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 12 Mar 12
Posts: 2895
Credit: 1,797,699
RAC: 396
United States
Message 1240710 - Posted: 3 Jun 2012, 17:30:30 UTC
Last modified: 3 Jun 2012, 17:33:14 UTC

You may be correct sir.
But you are speaking of credit per WU, I'm speaking of credit per TIME

If you look at my host, you will see that based on processing time of a unit alone, I'm getting 3x the credit.

simply put, 5000 seconds of processor time used to yield around 30 credits
and now 5000 seconds of processor time is yielding around 100.

My not Optimized WUs are disappearing from my list now, but every non Optimized WU that took time "y", was worth 1/3 the credit of optimized WU's taking time "y" to run...

My math cannot be wrong...?

I am new here and my understanding of credits is low. And I am NOT saying my computer is doing 3x as much work now, I'm simply saying I'm making 3x the credits per processor cycle, and on that I cannot figure the math in any different way.?
____________
-Dave #2

3.2.0-33

Profile Ex
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 12 Mar 12
Posts: 2895
Credit: 1,797,699
RAC: 396
United States
Message 1240721 - Posted: 3 Jun 2012, 17:41:58 UTC
Last modified: 3 Jun 2012, 17:47:43 UTC

And I'll add, in the last 24 hours, on my xeon box, I've already earned over 6,000 credits and that's NOT including work yet-to-be verified.

My normal RAC for this machine hovered around 2000, (at 35% proc usage)
I've doubled it's throttle to 70% before this adventure.
So with normal apps (at this throttle) I'd be looking at 4000 credits/day
But with optimized I'm getting over 6000 (very likely over 7,000)
This is at least a 50% improvement! and may be upwards of 75%...

In the end time will tell. I am taking daily logs of RAC, total credit, on a whole, and per-machine basis...
____________
-Dave #2

3.2.0-33

Profile perryjay
Volunteer tester
Avatar
Send message
Joined: 20 Aug 02
Posts: 3377
Credit: 16,211,348
RAC: 5,406
United States
Message 1240740 - Posted: 3 Jun 2012, 18:12:54 UTC - in response to Message 1240721.

My mistake, I thought you were comparing two work units of about the same time rather than the amount of credit you got in that amount of time. Lunatics does make a tremendous difference doesn't it? One other thing you should note though is that it takes about a month for RAC to settle in so you might see some strange numbers to begin with. Welcome to the big time! :)

____________


PROUD MEMBER OF Team Starfire World BOINC

Profile Ex
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 12 Mar 12
Posts: 2895
Credit: 1,797,699
RAC: 396
United States
Message 1240753 - Posted: 3 Jun 2012, 18:28:52 UTC

Lol, thanks!

Yea I'll be running my experiment for a month here, I expect my RAC to top 10,000. Then I'll cut out my extra machines, and go back to running just the xeon at 35%, at which time my RAC will level out around 4,000 or so.

I know it's not much. But I started with an RAC of 400 or so, just running one tiny virtual machine for seti.
____________
-Dave #2

3.2.0-33

Previous · 1 · 2 · 3 · 4 · 5 · Next

Message boards : Number crunching : My worthless non-GPU adventure

Copyright © 2014 University of California