FAH SMP Affinity Changer [Archive]

CyberRad

02-15-08, 09:09 AM

Hail my brother folders. :beer: I found this little gem on Xtreme Systems and didn't know if anyone has used it yet. I havn't seen it posted here yet.

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

Supposedly you can get a major increase with this program on the Core2 line of intel procs. Reports are coming in at about a 20% - 30% increase in PPD.

I am currently trying on 2 core2s and 1 quad. I'll let you know what I come up with ASAP.

NedClocker

02-15-08, 10:02 AM

Cool! Let us know!

CyberRad

02-15-08, 10:20 AM

So far it looks like I have gained any where from 1 to 2 minutes. That almost translates to about 1000PPD by my count. These results are after 2 hours of running this program. I'll update after I finish the units.

harlam357

02-15-08, 11:37 AM

Are you seeing the increase just on the quad or on both the duals as well?

Actually, someone posted this guy up months ago... but I don't think anyone here ever really put it through its paces. Please keep us updated on your findings.

dfonda

02-15-08, 04:56 PM

So if I have this straight it won't help if your running only one client...Correct?

Assigning affinities on QUAD if you run one SMP client does not make sense. from article.

Wicked_Pixie

02-15-08, 05:36 PM

I tried this last night just for kicks. It seems on par with running 2 SMP on native Linux or VMware.

Here's 2 WinSMP Q6600 @ 3200Mhz:
http://i10.photobucket.com/albums/a122/PixieMarj/2008-02-15_150953.png

Here's 2 SMP on Linux natively (Q6600 @ 3400Mhz):
http://i10.photobucket.com/albums/a122/PixieMarj/2xSMP-Linux3500Mhz-1.png

I might just stick with these... Vmware seems to slow down my rigs unless I use 4GB of RAM.

ihrsetrdr

02-15-08, 07:33 PM

NedClocker

02-15-08, 07:42 PM

3 minute check points. Wow.

I'm a little slow. What is it I'm supposed to see in those pics?

NedClocker

02-15-08, 07:44 PM

Wicked_Pixie,

It appears that your Ubuntu rig is running two 26xx WU's, thus not being detected as a quad by Stanford's assignment server.

You must have one of those 'stealth' motherboards that aren't recognized as running a quad...?

Which motherboards would those be? When I was running Win SMP on my quad, it was not getting assigned to the quad server. Of course, it's running 2 ubuntu VMs now.

ihrsetrdr

02-15-08, 07:53 PM

Which motherboards would those be? When I was running Win SMP on my quad, it was not getting assigned to the quad server. Of course, it's running 2 ubuntu VMs now.

I don't know, that's what I'm trying o find out. :D

I'm thinking of grabbing a Q6600 GO and one of those 'stealth boards'. :drool:

harlam357

02-15-08, 08:37 PM

Just installed on my E8400... we'll see if brings my frame times below 9:46... and they're running a little long right now since I've got Winamp cranking. :D

Wicked_Pixie

02-15-08, 09:02 PM

Wicked_Pixie,

It appears that your Ubuntu rig is running two 26xx WU's, thus not being detected as a quad by Stanford's assignment server.

You must have one of those 'stealth' motherboards that aren't recognized as running a quad...?

I don't know if it has anything to do with a particular board, but now that you have mentioned it, it could be. I have 2 GA-P965 DS3(rev. 3.3) running 2 SMPs each and all pick up either 2653 or 2605.

I also experimented a bit. :shy: First, I had a C2D in it. I let it finish a WU. Then start a new folder and let that finish a WU. Then I swap the C2D with a quad and start both SMPs. So far it's consistently picking up 26xx for about a week now.

Forgot to mention earlier, the first few percentile takes a bit longer...

I don't know if this is a fluke or not. I may try this on an Abit IP35-E in a week or so... I got tired swapping parts at the moment. Maybe someone can confirm it with a different board...

3 minute check points. Wow.

I'm a little slow. What is it I'm supposed to see in those pics?

Oh, I was only pointing out that the affinity changer works, in comparison to running 2 SMP Lionux. Of course time will tell as I have yet to finish my first set of Wus. Anyway, more options for us running 2x SMPs. :attn:

But I also read that Stanford prefers having the WU's returned quickly and kinda frowns on 2x SMP on quads... :confused:

I don't know, but how different is it from someone who runs SMP but games a lot or can only run it for several hours a day?

harlam357

02-15-08, 09:23 PM

Yes sir... 9:22 with Winamp running... should pick up another few seconds while the machine is idle. :D

ihrsetrdr

02-15-08, 09:29 PM

I don't know if it has anything to do with a particular board, but now that you have mentioned it, it could be. I have 2 GA-P965 DS3(rev. 3.3) running 2 SMPs each and all pick up either 2653 or 2605.

I also experimented a bit. :shy: First, I had a C2D in it. I let it finish a WU. Then start a new folder and let that finish a WU. Then I swap the C2D with a quad and start both SMPs. So far it's consistently picking up 26xx for about a week now.

Forgot to mention earlier, the first few percentile takes a bit longer...

I don't know if this is a fluke or not. I may try this on an Abit IP35-E in a week or so... I got tired swapping parts at the moment. Maybe someone can confirm it with a different board...

Hmmm, I have at least one, if not a couple GA-P965 DS3(rev. 3.3) motherboards running quads; I bet your boards having previously had dual core processors in them has something to do with it. Hmmm. Did you keep the original operating system, from when the dual cores were in the DS3 rev.3.3's?

But I also read that Stanford prefers having the WU's returned quickly and kinda frowns on 2x SMP on quads... :confused:

Well, Stanford may frown on doing that, but they surely realize that many quad owners are buying this hardware exclusively to fold on, and as such must have an appreciation for this level of contribution.

If it weren't for folding, my electirc bill would definately be a couple hundred dollars a month less. ;)

harlam357

02-15-08, 10:13 PM

But I also read that Stanford prefers having the WU's returned quickly and kinda frowns on 2x SMP on quads... :confused:

My 2xVM SMP Quad returns WUs more quickly than any of my AMD dualies... so in my eyes they're getting their work done quicker. :D

ChasR

02-15-08, 11:07 PM

On p2653
Based on Wicked Pixie's numbers.
2 x Win SMP with affinity changer = 1220 ppd/GHz
2 x native Linux = 1376 ppd/GHz

Wicked #s might be a little better on native Linux if checkpoints were set to 15 min.

harlam357

02-15-08, 11:48 PM

Warning to those running WinSMP on AMD dualies.

The SMP Affinity Changer appears to lower the ppd on the box I tested it on (939 Opteron).

I'll test on other rigs but this is my initial result. :shrug: I would expect that it doesn't benefit like C2 because of the non-shared L2 structure. Phenom with shared L3 may be a different story.

Adak

02-16-08, 12:34 AM

I've installed the SMP Affinity Changer in my C2D (E6700), and noticed:

1) When I run other programs, the balancing of the thread cores by AC, throws production into a cocked hat - folding really slows down. However, it is not noticeably slower than it was when these same programs were run concurrent with folding, before AC was installed.

So no difference, there.

2) When I am off the computer, or just doing light stuff with it (browsing, etc.), then folding times have decreased and ppd have increased noticeably. FahMon was reporting 1585 ppd for this rig (at stock speeds), before AC, and now reports 1720 ppd.

That's a sizeable production increase for a C2D rig (note Harlam's post above on AMD's dual core problems with AC, above).

If you have a C2D folder, I recommend immediately installing Win SMP Affinity Changer. It has caused no harmful effects whatsoever, doesn't communicate to anybody/anyplace on the net, and is so small and lean on resources, it almost is undetectable.

The url for it:
http://distributed.org.ua/forum/index.php?showtopic=1149

Note:
On my browser (firefox), I had no OK button to d/l the program when I tried. I had to click a few times on the grayed out OK button in the d/l pop-up window, to get it to show the OK button, correctly, and allow the d/l to happen.

ChasR

02-16-08, 07:55 AM

For hrsetrdr,
My boards that "avoid" the "quad only server:
ASUS Commando
ASUS P5B (vanilla)
ASUS P5K (vanilla)

Boards that dont:
Gigabyte DS3R
ASUS P5K Premium

I'll check and see what HayesK is running as all of his hit the "quad only" server. Darnold 87 has one that avoids it, so I'll check on that as well.

I have no idea why some hit the quad server exclusively and some never do. It could be bios related. The logs on all of them report 4 cores detected.

For all running Intel C2 rigs,

Affinity changer is a great little program. If you don't want to run in a vm, you can still come close in ppd output. In my experience however, the VM approach has the advantage of not stopping on the loss of the network making it more suitable for wireless networks or otherwise flakey networks

wingless

02-16-08, 08:18 AM

harlam357

02-16-08, 10:24 AM

Ok, now I've got idle numbers for AC vs no-AC on my E8400.

AC: 8:53tpf = 2810ppd

no-AC: 9:46tpf = 2595ppd

Net gain: 215ppd :attn:

I agree with Adak... this program is totally safe. It's a lightweight service written in .NET and is very simple.

I just ran reflection against the Affinity Changer code, basically a disassembly, and I see nothing of concern. In no way is it talking to the outside world and/or doing anything malicious to your system.

This one's a keeper boys... :beer:

LandShark

02-16-08, 10:44 AM

I don't get it....

so, this little program helps increase your ppd on C2 processor right? but in what situation?? in running 2xWinSMP on a quad that's getting non quad only WU?? or on a single WinSMP situation (on quad? or dual? or both?)??

harlam357

02-16-08, 11:52 AM

but in what situation?? in running 2xWinSMP on a quad that's getting non quad only WU?? or on a single WinSMP situation (on quad? or dual? or both?)??

- 2xWinSMP on a quad (non-quad units)

- 1xWinSMP on a dual

...using it on a quad with 1xWinSMP appears to provide no gain.

LandShark

02-16-08, 12:45 PM

- 2xWinSMP on a quad (non-quad units)

- 1xWinSMP on a dual

...using it on a quad with 1xWinSMP appears to provide no gain.

so, 1xWinSMP on a dual/C2D helps. but not on a Quad/C2Q that's getting a quad only WUs nor better than 2xVMware/LinuxSMP on dedicated folder.

will gonna try it on C2D running WinSMP then. ;)

Wicked_Pixie

02-16-08, 05:40 PM

Hmmm, I have at least one, if not a couple GA-P965 DS3(rev. 3.3) motherboards running quads; I bet your boards having previously had dual core processors in them has something to do with it. Hmmm. Did you keep the original operating system, from when the dual cores were in the DS3 rev.3.3's?

I'd like to think swapping CPUs had something to do with it. Otherwise all that work was a waste of time. :bang head
I kept the original OS. On your P965 DS3's, do you get quad WU's exclusively too?

My 2xVM SMP Quad returns WUs more quickly than any of my AMD dualies... so in my eyes they're getting their work done quicker. :D

Well, Stanford may frown on doing that, but they surely realize that many quad owners are buying this hardware exclusively to fold on, and as such must have an appreciation for this level of contribution.

If it weren't for folding, my electirc bill would definately be a couple hundred dollars a month less. ;)

Ack! Don't remind me of electric bills. I want to :cry: every time I see it.

The way I see it, running SMP the proper way yields about 2 WUs every 22-25 hrs. depending on the rig. With 2x SMP, I can shave off ~3hrs or so. That's roughly 2 additional WU's a week for each quad rig... for the same amount of electricity.

So yeah, I am now sold on running 2xSMP.:)

Wicked_Pixie

02-16-08, 05:42 PM

Wicked #s might be a little better on native Linux if checkpoints were set to 15 min.

I will try that. Thank you. :)

LandShark

02-16-08, 08:55 PM

as some of you know, my farm is almost all favorite of DS3 (P965, P35, all revision), and I can tell you, this is not the "secret board". at least not in my case!

ok, this SMP Affinity Changer just got the shark approval!!

http://landshark-racing.com/pics/lager1.jpg

the average ppd gain on my C2Ds is about 150-180ppd!!

here's some number for you before and after....

Project : 2653
Core : Gromacs SMP
Frames : 100
Credit : 1760

-- station1 --
Min. Time / Frame : 13mn 13s - 1917.58 ppd
Avg. Time / Frame : 14mn 32s - 1743.85 ppd

-- station5 --
Min. Time / Frame : 13mn 09s - 1927.30 ppd
Avg. Time / Frame : 14mn 34s - 1739.86 ppd

-- station7 --
Min. Time / Frame : 14mn 18s - 1772.31 ppd
Avg. Time / Frame : 16mn 10s - 1567.67 ppd

-- station4 --
Min. Time / Frame : 12mn 54s - 1964.65 ppd
Avg. Time / Frame : 13mn 40s - 1854.44 ppd

I'm going to install this on the rest of my C2D WinSMP rigs!

McTimson

02-17-08, 11:16 AM

Wow...just went from 15:25/frame to 13:53/frame on my C2D. This should be required for any Windows C2D machines, it's a great little boost.

Adak

02-17-08, 01:43 PM

Thank you, CyberRad! Great tip for Intel multi-core rigs.

ihrsetrdr

02-17-08, 02:01 PM

For hrsetrdr,
My boards that "avoid" the "quad only server:
ASUS Commando
ASUS P5B (vanilla)
ASUS P5K (vanilla)

Boards that dont:
Gigabyte DS3R
ASUS P5K Premium

Thanks ChasR. It's still a 'crap-shoot' though, maybe BIOS, maybe related to prior occupancy by dual core. Of the boards you listed, had any originally had dual cores in them?

*slightly off topic* a curious thing: on one of my quads I put W2K to get VMware up & running quickly(i like that) but had forgotten that W2K only 'sees' 2 processors(as evidenced by Task Manager). However, the WUs running in VM on the W2K rig show comparable TPF to the P26xxs running on the other rigs that have XP as the Host O/S. Odd.

I'd like to think swapping CPUs had something to do with it. Otherwise all that work was a waste of time. :bang head
I kept the original OS. On your P965 DS3's, do you get quad WU's exclusively too?

Yes, all my quads are DS3's, and all used to get the quad-only WU's; this past week I've conververted all of the rigs to x2 VMs, each running one Ubuntu and one notfred's folding image (http://reilly.homeip.net/folding/).

Ack! Don't remind me of electric bills. I want to :cry: every time I see it.

Ya, here in da schtate of Calyfornya ve pay da highest electric rates in da vorld!

http://www.geocities.com/tthomas_sbsd.empl/Forum_Images/Ahnold2.jpg

LandShark

02-17-08, 03:30 PM

Ack! Don't remind me of electric bills. I want to :cry: every time I see it.

oh, and NEVER remind me the power bills too!!! http://www.ocforums.com/showthread.php?t=518318&highlight=power+bill+price+point

Wicked_Pixie

02-17-08, 10:33 PM

Ya, here in da schtate of Calyfornya ve pay da highest electric rates in da vorld!

[/QUOTE]
Rofl on the Governator impression. I get a 5-week reprieve on my power bill soon.
My boss asked me to house sit his home soon and I asked if I can bring a couple of PC's with me. I think I'll just bring them all. Shhhhhh, don't tell.

oh, and NEVER remind me the power bills too!!! http://www.ocforums.com/showthread.php?t=518318&highlight=power+bill+price+point

:eek: but... but... ihrsetrdr started it. :P

Yeah, tough choice for me. Sweat to death (blazing Cali. summer) and fold or be comfy for the whole summer but not able to help. Hmmm, I might just scale mine down to a couple. We'll see.

CyberRad, thanks a bunch for sharing. It's been helpful, especially on my C2D. :thup:

xsuperbgx

02-17-08, 10:54 PM

Which is better? this and winsmp or smp on ubuntu in vmware ?

ChasR

02-18-08, 06:50 AM

Affnity changer gets Win SMP close to VM/Linux and native Linux.
On p2653 on an E6600:

Win SMP w/AC = 660 ppd/GHz
VM/Linux = 695 ppd/GHz
Linux = 720 ppd/GHz

Quad comparisson:
2 x Win SMP with affinity changer = 1220 ppd/GHz
2 x VM/Linux = 1335 ppd/Ghz
2 x native Linux = 1396 ppd/GHz

Linux SMP also has the advantage of not stopping when you lose network connectivity.

ChasR

02-18-08, 06:57 AM

Of my magic boards, only the P5B underwent a cpu swap after the SMP install. I should be doing another swap today and we'll see what happens.

CyberRad

02-18-08, 07:00 AM

Sorry about the lack of response. Been working on on the drag car. March is creeping up quick.

Anyways... I found little to no change in my quad. I'll try running the 2XSMP clients. As far as the C2s, my E4300 saw a few second change and the E6600 saw no change. I might be hurting myself at night when the machines do their maintenance (machines not only used for folding).

CyberRad

02-18-08, 07:30 AM

I just checked the stats Wicked and damn... your going to barrel through me soon. You've only been folding for 2 months now and your about to pass. LOL

Anyways no problem for the program. I saw it and thought it would help us get to that number 2 spot a little faster.

ihrsetrdr

02-24-08, 11:59 PM

Of my magic boards, only the P5B underwent a cpu swap after the SMP install. I should be doing another swap today and we'll see what happens.

ChasR,
Did you get a chance to do the cpu swap? if so, how'd it go? I'm coming into some 'built-in'(mandatory) overtime at work, for a couple months; this of course means...time to build another machine! :D Maybe I'll snag one of those magic boards to tide me over, until the 45nm procs come out. I have an "old" dual core that's available to try this magic board setup...

Doggonit...I can see the size of my farm starting to creeping back up(along with the electric bill). :eek:

ChasR

02-25-08, 07:26 AM

Tim,
I did the cpu swap and the first two WUs downloaded were p2653s. I thought I had it figured out. However the next two were p3043 and p3060, then back to p2653, then back to the same p3043 that died @ 8%. I'll check it again when I get to work.

ChasR

02-25-08, 10:34 AM

This P5K/Q6600 appears to be struck on the Quad only server.

ihrsetrdr

02-25-08, 11:48 PM

This P5K/Q6600 appears to be struck on the Quad only server.

Oh well, I'd be willing to bet that the 'magic' board's days would have been numbered anyway, as Kasson & Co.continued to refine the client.

A lot of talk around here about the Wolfdale...might consider a Wolfy to tide me over, until the Q9450's come out.

Time just crawls when I have a pocket-full of overtime money. :D

Threader

03-02-08, 07:19 PM

Can someone explain this quote from the Affinity Changer webpage

"For quad core processors, first SMP client is assigned to 1st and 3rd core, 2nd SMP client assigned to 2nd and 4th core."

WTF?

I have been manually running my affinities "1&2" & "3&4" for my Ubuntu instances in VMWARE.

Don't tell me I've been running an inefficient core affinity pair for the last 6 months?

Excuse me while I quietly slip into my tub and slash my wrist in anguish :bang head

ihrsetrdr

03-02-08, 07:38 PM

Shiggity

04-07-08, 05:04 PM

Bump, can someone explain to me what this actually does?

I'm very interested in getting it for my 1xSMP dual core setup, but when I'm folding I have tons of apps running usually, and someone said that would kill your PPD.

Can someone clarify all this affinity changer business for me. I really appreciate it, maybe a guide? :)

ChasR

04-07-08, 05:17 PM

As I understand it, Affinity Changer is managing cpu usage without regard to cache. It switches FAH threads from cpu core to core depending on the load on each cpu core to maximize cpu efficiency. In VMware, without the ability to switch, you're trying to ensure the cpu isn't requesting data on the cache not controled by it to maximize cache efficiency.

surfasb

05-18-08, 01:48 AM

harlam357

05-18-08, 09:50 AM

To maximize a 2xVMWare/Linux/SMP setup on a quad go to the Windows Task Manager and find each vmware-vmx.exe process.

Right-Click -> Set Affinity -> choose Cores 0 & 1 for one process and Cores 2 & 3 for the second. You can also set each process priority to Low as well.

This will do two things, keep each vmware process on it's own cpu die which stops any cache swapping across the fsb between the two dice.... and by setting priority Low the machine will yield processor cycles to other normal applications when using the machine, much like it would using a normal FAH install, SMP or standard client, they all run at Low process priority.

The version of VMWare should not matter... but, I'm using v1.0.4 on one rig and v1.0.5 on two others.