Affinity Changer Update

[alan2308]

Saw this over at OCF. AC 1.0.5 now supports dual quads.

http://distributed.org.ua/forum/index.php?showtopic=1149

 

[APOLLO]

Thanks for the info, Alan. I was posting there earlier today helping out, I don't know how I missed it. I'll give it a spin.

 

[Pocatello]

Dual Quads for the Win!

 

[APOLLO]

Dual Quads for the Win!

Not really. They aren't as good as two separate single quad systems for various reasons, not least of which is the bad memory subsystem. The relatively poor OC factor is another. Excessive RAM heat and extremely limited HSF choice is yet another. They only advantage dual quads offer for F@H is space savings, nothing else. One would be lucky to see 6000PPD, I get a lot less.

So far, I haven't seen anything advantageous with the new AC, unless I haven't configured the optimum setup yet. It seems slower to have this running than my previous configuration without it. I will need to try some changes before giving my final judgment. I just wish the author was a little more informative on how to run this optimally with 8 cores.

 

[MixManSC]

I'll try this on the box in my sig in an hour or so when I get to my shop. Thats a dual quad core xeon running 2 instances of the SMP console client. That box has never gotten anywhere near 6K PPD. I wonder if it AC would allow it to now run 4 SMP clients instead of just 2?

Edit - Just checked. The dual quad box I have has been averaging right at 3000PPD running 2 smp clients chewing on a steady diet of sucky 2665 work units. I'll update in a few hours once I get some firm numbers going again. I just installed the new AC. After a few hours I'll be able to see if there is a difference. Then I'm going to try and launch 4 smp clients and see what that does. Any increase is better than none though so I'll take what I can get.

Edit 2 - Okay after about 4 steps running 2 smp clients it looks like zero performance gain. Here's where it gets fun. Previously there was no way to get 4 win smp clients running properly on a dual quad core box. I now have 4 win smp clients all running at the same time and all working on different work units. The 2 original clients are continuing on the 2665 work units they were in progress on. The 2 new clients, 1 got a 2665 and the other got a 2653. I'll update soon with new performance numbers.

 

[Crazy Larry]

MixManSC,

Sounds like it starting to get interesting with that dual quad...

 

[MixManSC]

We shall see - I need to get a few hours of data with all four instances running to really get a good grip on what the numbers will work out to be. The first frame is completing on all 4 right now and it's all over the place so far. One shows 30 minutes, one shows 40, and one shows 36, I staggered starting them so the fourth has not yet completed its first frame. Really need several frames each to get a better idea.

 

[MixManSC]

So far not too promising. It's looking like the frame times are pretty much double and the PPD is about the exact same. Only difference is it's now crunching 4 units at once instead of 2. There is a distinct disadvantage as well - now it may not always make the "preferred" deadlines. I'll run it like this for the rest of the day and see how things looks once a substantial number of frames have completed but so far I'm not too thrilled. I so far agree fully with Apollo's comments. If there was some information as to how this would be best configured it would help greatly.

 

[MixManSC]

Verdict for day 1. Sucks. I can now run 4 win smp clients but ppd is about the exact same. Going to let it run through the work it's on and stop the 2 extra clients.

As of now it's all over the place on the work units. One 2665 is getting average 43:36 per frame, next 2665 is getting 26:43, next 2665 is getting 39:35, and the 2653 work unit is getting 36:14. Shows a total PPD on the box as 2912 - pretty much exactly what I was getting with 2 instances crunching 2665 work units. Oh well - was worth a try to see if it would help. Until someone comes up with a way to configure it to give some sort of actual improvement I'm probably going to stop AC on the dual quad box.

 

[APOLLO]

Verdict for day 1. Sucks. I can now run 4 win smp clients but ppd is about the exact same. Going to let it run through the work it's on and stop the 2 extra clients.

As of now it's all over the place on the work units. One 2665 is getting average 43:36 per frame, next 2665 is getting 26:43, next 2665 is getting 39:35, and the 2653 work unit is getting 36:14. Shows a total PPD on the box as 2912 - pretty much exactly what I was getting with 2 instances crunching 2665 work units. Oh well - was worth a try to see if it would help. Until someone comes up with a way to configure it to give some sort of actual improvement I'm probably going to stop AC on the dual quad box.

What's the specs of your system? Does it see heavy use? Seems kind of low level of production for a dual quad, with or without AC.

When I read about the coming release of AC with support for 8 cores, I was quite enthusiastic and impatient. I tried the original AC and it didn't grant me any advantage, but nether does this version. I tested the new version on a dual Clovertown and dual Harpertown. I noticed no significant improvement running with it or without it.

If this utility works like the original, we need to install 4 clients in order to process two threads per core. However, 4 clients bogs things down so much with the predominant 2665 WUs that frame times become huge as a result, and the completion dates approach close to the deadlines. Not good.

In conclusion, either the author didn't have a dual quad on hand to test his utility, or he did something very wrong and badly needs to go back to the drawing board. For those that own these systems, the best one can do to maximize production is to install VMs. If that's not an option, I can vouch for running 3 clients instead of 2 or 4. Yes, it's an odd number but it works for me. I have tried everything from 2 to 5 clients back when 2653 WUs dominated, and by far the best setup was 3 clients on a dual quad running Windows. They also finish well before the deadlines.

The tri-client setup earned me an average of 1000PPD (sometimes more, YMMV) advantage over dual clients with the 2653, but that WU is a notoriously poor scaler. Ex. a 2.5GHz Clovertown with 2 clients = 11-12 min/frame; same system with 3 clients = 13-14 min/frame. A 2.8GHz dual quad Harpertown was producing nearly 6200PPD with 3 Windows clients and 3x2653 WUs – no VMs. The new WUs still garner a noticeable improvement running 3 clients over 2, so this setup remains until things change, maybe with the advent of the A2 core. We'll shall see.

 

[MixManSC]

It's dual Clovertowns - box in my sig. I guess maybe tomorrow when I'm watching it, I'll just kill 1 client and take it down to 3 instead of 4. See how that works out.

 

[APOLLO]

It's dual Clovertowns - box in my sig. I guess maybe tomorrow when I'm watching it, I'll just kill 1 client and take it down to 3 instead of 4. See how that works out.

The closest system I have to the one in your sig is a dual Clovertown @2490MHz and 2GB RAM running Win XP SP2. It is currently folding 3 clients each with P2665 WUs. FahSpy is reporting 4058 PPD for this system and 3 client setup.

Another thing about the tri-client setup is that each client processes its WU in approximately the same amount of time per frame as the other clients, providing the WUs are of the same type. IOW, there's consistency that we see with the dual client setup but three clients produce more PPD. Running a number of clients above three causes wild performance disparities between clients and less PPD, with the aforementioned very long frame times. Here's a short summary on the above system's current WU production:

Client 1:

[23:19:12] Project: 2665 (Run 1, Clone 167, Gen 7)
[23:19:12]
[23:19:15] Entering M.D.
[23:19:23] Calling FAH init
[23:19:25] Read topology
[23:19:25] (Starting from checkpoint)
[23:19:26] out of 250000 steps (5 percent)
[23:19:26] ater
[23:19:26] Writing local files
[23:19:26] Completed 14389 out of 250000 steps (5 percent)
[23:19:38] Extra SSE boost OK.
[23:24:31] Writing local files
[23:24:32] Completed 15000 out of 250000 steps (6 percent)
[23:44:24] Writing local files
[23:44:24] Completed 17500 out of 250000 steps (7 percent)

PPH (points per hour): 57.89
PPD (points per day): 1389
Avg time per step: 0:19:54

Client 2:

[23:18:21] Project: 2665 (Run 3, Clone 530, Gen 9)
[23:18:21]
[23:18:30] Entering M.D.
[23:18:38] Calling FAH init
[23:18:40] Read topology
[23:18:40] ocal files
[23:18:40] rom checkpoint)
[23:18:40] Read checkpoint
[23:18:40] Protein: HGG in water
[23:18:40] Writing local files
[23:19:02] Extra SSE boost OK.
[23:19:02] Writing local files
[23:19:03] Completed 0 out of 250000 steps (0 percent)
[23:39:27] Writing local files
[23:39:27] Completed 2500 out of 250000 steps (1 percent)
[23:59:51] Writing local files
[23:59:51] Completed 5000 out of 250000 steps (2 percent)

PPH (points per hour): 56.49
PPD (points per day): 1356
Avg time per step: 0:20:23

Client 3:

[23:17:53] Project: 2665 (Run 2, Clone 808, Gen 12)
[23:17:53]
[23:18:02] Entering M.D.
[23:18:08] Calling FAH init
[23:18:10] Read topology
[23:18:10] (Starting from checkpoint)
[23:18:11] Read checkpoint
[23:18:11] Protein: HGG with glycosylations
[23:18:11] Writing local files
[23:18:23] Extra SSE boost OK.
[23:18:24] Writing local files
[23:18:25] Completed 0 out of 250000 steps (0 percent)
[23:39:29] Writing local files
[23:39:29] Completed 2500 out of 250000 steps (1 percent)
[00:00:30] Writing local files
[00:00:30] Completed 5000 out of 250000 steps (2 percent)

PPH (points per hour): 54.73
PPD (points per day): 1313
Avg time per step: 0:21:03

If you have any questions regarding dual S771 systems, let me know. I can post specs and client statistics for my other systems as well.

 

[MixManSC]

Okay - took it down to 3 smp clients. The new AC is still running as well. Keep in mind I'm running native win smp stuff - no vmware here. 3 clients all on 2665, 33:42, 35:52, and 25:42 per frame. Now I'm getting worse overall PPD than I was with 2 clients and no affinity changer. Was getting right about 3000 PPD with 2 2665 units going. Now showing about 2700 PPD. Looks like I'm going back to 2 clients and no AC.

 

[Tigerbiten]

I'll running two Dual X5455's systems @2.66 Ghz.
Both are running quad VM's.
Ones pulling in 6,400 PpD with 4x p2605.
The other ones pulling in 5,600 PpD with 3x p2605 & 1x p2665.
The p2665 is slowing it down by ~400 PpD.

I'm still waiting for the second CPU and the Ram for my skulltrail system.

Luck ................

 

[MixManSC]

I guess ultimately virtual machines is the way to go. I'm going to have to bite the bullet and eventually figure out how to use vmware. I've never once used it, ever, and also dont know crap about using any flavor of linux so it's quite intimidating.

 

[APOLLO]

Nice production there Tiger, but this thread solely concerns native Windows clients with AC, not multiple VMs which are duly served by numerous other threads. Everyone knows that VMs result in greater overall production. Nevertheless, I had one of my systems produce very close to your figure (6200PPD) without VMs, no 64-bit Linux, no AC and only 3 Windows clients.

If Stanford was in the habit of releasing 2653 WUs in the same quantity as a few months ago, I would still see that level of production. How much would you wager 6200PPD can be bumped up running a working version of 8-core compatible AC?

 

[MixManSC]

One question Apollo - I thought without AC no one could get more than 2 win smp clients to run right on a dual quad? Some trick to it or does it just work now?

 

[APOLLO]

One question Apollo - I thought without AC no one could get more than 2 win smp clients to run right on a dual quad? Some trick to it or does it just work now?

If I was able to see your system in person, I could probably make a dozen SMP clients run concurrently. Seriously, MixManSC it always worked for me and I had up to 5 clients running briefly once to test the best native Windows setup for 8 cores. My results proved time and again that 8 Intel processor cores ran optimally under Windows XP with 12 SMP cores. I can't say for sure why.

Your system can definitely stand to see an improvement. There's no reason why my system being only ~160MHz faster should have over 1000PPD advantage over yours working on the same WUs. Don't you think that's strange? Maybe you should uninstall AC (until it is improved) and see then if additional clients work. I have tried this on several systems and they all demonstrated an improvement increasing from 2 to 3 clients. They are dedicated folders, though.

 

[MixManSC]

Thanks man - I stopped AC and set it's service to manual for now. See if the 3 running clients pick up some steam. While that box does have 16GB in it Win only have 4 available to it - 12 is configured as a ramdrive. During the day the machine does almost always have Photoshop CS3, FlexiSign Pro and Production Manager running along with Illustrator. I try to close all the other apps down when I'm not actively using them though.

 

[APOLLO]

Thanks man - I stopped AC and set it's service to manual for now. See if the 3 running clients pick up some steam. While that box does have 16GB in it Win only have 4 available to it - 12 is configured as a ramdrive. During the day the machine does almost always have Photoshop CS3, FlexiSign Pro and Production Manager running along with Illustrator. I try to close all the other apps down when I'm not actively using them though.

Maybe it's hard to believe but here's a screenshot of three clients with their frame times. Each client is working on a P2665 WU. This is the system I mentioned above, dual Clovertown @2490MHz with 2GB of RAM running Win XP-32 SP2.

 

[Tigerbiten]

Stats from FahInfo.org.
p2665 runs at ~160 PpDpGhz on a Quad Kentsfeild and at ~220 PpDpGhz on a Dual Conroe.
So 160 PpDpGhz x 2.33 Ghz x 8 Cores = ~3,000 PpD.
MixManSC may be around the right numbers for twin clients.
Running 3 copies should give you closer to dual speeds without overloading the memory bus to badly.
So 200 PpDpGhz x 2.50 Ghz x 8 Cores = ~4,000 PpD.
Which explains the 1,000 PpD difference from only a 160 Mhz increase in speed
Running 4 copies probably overloads the memory bus so doesn't give you any real advantage over 3.

That one reason I'm interested on how well the skulltrails is going to fold.
Load it up on a DDR2-800 FB Dimms to run at the 1600 FSB and the memory throughput will that bit more.

Luck .............

 

[MixManSC]

Oh I believe you - I just wish I could even get 4K ppd out of this box. Maybe it's my client configs.... I see you have some switches like local and forceasm. I have no command line switches set at the shortcut level, the clients are all 5.92 using the newer deino mpi, all configured as defaults except machine id is specific to each, allow wu over 5mb is on, and advmethods is on in the config. Anything you can think of? Now that I've stopped AC you think I should stop and restart the clients?

Thanks again for the help and posts you have made on this.

 

[APOLLO]

I tried with and without the -local switches and it didn't make a difference. I always have the -forceasm switch by habit, but Tiger can better tell you if that makes any real difference at all with SMP. The -verbosity switch gives you more information on what the client is doing. I have only the bare minimum required services in this Windows installation, maybe that helps a bit with the performance. Everything is stripped down including the GUI to reduce resource consumption, but how much can that add to F@H production?

 

[Xilikon]

Oh I believe you - I just wish I could even get 4K ppd out of this box. Maybe it's my client configs.... I see you have some switches like local and forceasm. I have no command line switches set at the shortcut level, the clients are all 5.92 using the newer deino mpi, all configured as defaults except machine id is specific to each, allow wu over 5mb is on, and advmethods is on in the config. Anything you can think of? Now that I've stopped AC you think I should stop and restart the clients?

Thanks again for the help and posts you have made on this.

Try 3x 5.91 and see. The 5.92 is a bit too early to be considered beta and is not that efficient.