Fah Smp Afiinity Changer - It Works!

I hear you Killer, but it does zero for me. Won't even set affinity.
Two different systems, Vista 64 and Vista x86. Nothing.

I gotta call shens.
 
You've gotta make sure the service is running and you install it on admin ;)
 
Killer[MoB];1032156589 said:
I have 3 systems I am running it on. Two of them are dual core intel chips. On my E6600 it is 9.4% faster after installing affinity changer. On my E4500, I don't know the exact times but it is faster after installing affinity changer.

The third system is a quad. As pointed out before, it does nothing on a quad running one instance of WinSMP. I have 2 instances of WinSMP running using affinity changer. I'm getting frame times of 11:12 to 11:32. So depending on how much use the machine gets, PPD can be as high as 4394 to 4524 PPD. Running just one instance, I was getting frame times of around 8:10 which is 3103 PPD for the machine.

Is that proof. It's enough for me. Plus there are posts going back several weeks on this forum and back to last year on other forums to back up that it does work.
Plus, before the little program was made, you could manually set affinity with the same results as long as you set the right fahcores to the right cpu cores. That's just a pain though. AC does the work for you.
Click to expand...
I'm not doubting that AC works but what I wish to know is why I have not seen any benefit whatsoever on dual quads? If it works on duals and quads based on the core architecture, why not dual quads?
 
I guess it's just not written to do anything on a dual quad just like it doesn't even try to set affinity on a quad with only one instance running. It probably doesn't know what to to do since CPUs 4-7 are present. Would be a good question for the developer of the proggy.
 
I heard that AC actually downloads some special brand of Ukrainian Gay Gymnast Amputee pr0n.
 
APOLLO said:
I'm not doubting that AC works but what I wish to know is why I have not seen any benefit whatsoever on dual quads? If it works on duals and quads based on the core architecture, why not dual quads?
Click to expand...

Well actually, there might be an easy justification of that. If you tried running 3 instances of WinSMP on a dual quad, you might see some benefit.

This is all speculation on my part since I haven't really studied core architecture in depth.

The reason I believe it works so well on dual cores is because its rather costly for a process to jump from one core to another (at least in Core architecture). Since WinSMP spawns 4 core processes, there's 2 more processes than there are available cores. This imbalance causes the processes to contend for CPU time and forces the OS to continually assign and reassign processes to each core. As one process jumps to another core, its data in the cache has to be moved there as well. And this is a rather costly operation.
The affinity changer prevents this reassigning and keeps each process on a specific core.

This is why there's not much benefit for those running one instance of WinSMP on a quad core. Each WinSMP core process has one physical CPU core that can handle it. Thus, the OS (Windows) doesn't need to do any reassigning of processes to different cores since there's no contention over CPU time.
But when you're running two instances of WinSMP, you have 8 core processes and only 4 CPU cores and once again an imbalance. That's why affinity changer helps.

This is why on a dual quad (8 CPU cores) you might only see benefits when you're running at least 3 instances of WinSMP (12 core processes).
 
relic said:
Tried to get affinity changer to do anything, but it does absolutlely nothing.
My personal vote is that it's completely worthless.
Click to expand...

Sorry to hear you haven't had any luck with this affinity deal, I can honestly say using a C2D E6600 that's mildly OC'ed to 2.8 GHz, using WinXP /w SP2 that I have got a 2x mins per frame improvement. I've included part of my F@H log file "after" the affinity program, but I don't have a "before" the infinity program to post. Sorry :(

21:33:45] Folding@Home Gromacs SMP Core
[21:33:45] Version 1.74 (March 10, 2007)
[21:33:45]
[21:33:45] Preparing to commence simulation
[21:33:45] - Ensuring status. Please wait.
[21:34:02] - Looking at optimizations...
[21:34:02] - Working with standard loops on this execution.
[21:34:02] Examination of work files indicates 8 consecutive improper terminations of core.
[21:34:02] es status OK
[21:34:08] 57862 -> 15216508 (decompressed 514.4 percent)
[21:34:08] 4.4 percent)
[21:34:09] 3 (Run 33, Clone 56, Gen 70)
[21:34:09]
[21:34:09] 56, Gen 70)
[21:34:09]
[21:34:12]
[21:34:12] ntering M.D.
[21:34:18] nit
[21:34:19] ein: Protein in POPC
[21:34:19] WritingCompleted 310Completed 310000 out of 500000 steps (62 percent)
[21:34:19] cal files
[21:34:19] Completed 310000 out of 500000 steps (62 percent)
[21:34:20] Extra SSE boost OK.
[21:48:16] Writing local files
[21:48:16] Completed 315000 out of 500000 steps (63 percent)
[22:02:17] Writing local files
[22:02:17] Completed 320000 out of 500000 steps (64 percent)
[22:16:43] Writing local files
[22:16:43] Completed 325000 out of 500000 steps (65 percent)
[22:30:49] Writing local files
[22:30:49] Completed 330000 out of 500000 steps (66 percent)
[22:44:33] Writing local files
[22:44:33] Completed 335000 out of 500000 steps (67 percent)

I think this is a 14 mins frame time, you'll have to believe the before frame time was 16 mins. You really don't have to believe it, it's not that important :rolleyes: , but I know this SMP infinity program works on a C2D E6600 with WinXP /w SP2. (all my other boxen use Linux, 3x Ubuntu flavor, 1x Mandriva 2008 flavor) :D

 
CoW]8(0);1032157555 said:
Well actually, there might be an easy justification of that. If you tried running 3 instances of WinSMP on a dual quad, you might see some benefit. This is all speculation on my part since I haven't really studied core architecture in depth.

[...]

This is why there's not much benefit for those running one instance of WinSMP on a quad core. Each WinSMP core process has one physical CPU core that can handle it. Thus, the OS (Windows) doesn't need to do any reassigning of processes to different cores since there's no contention over CPU time.
But when you're running two instances of WinSMP, you have 8 core processes and only 4 CPU cores and once again an imbalance. That's why affinity changer helps.

This is why on a dual quad (8 CPU cores) you might only see benefits when you're running at least 3 instances of WinSMP (12 core processes).
Click to expand...
Thanks for the explanation. I tested at least 3 instances on a dual quad. I tried every combination from 3-5 instances and compared with and without the AC, but unfortunately there wasn't any perceptible change in PPD. I figure the program is limited to optimizing F@H processes up to a maximum of four CPU cores. If that's the case, it shouldn't be to difficult for the person who wrote the program to expand its function for optimizing affinity up to 8 cores. I guess he would have to see a big enough need first before bothering.
 
Quoted from AC site:

How does it work?

After some time interval (10 mins) utility checks, if there are 4 processes of FahCore_a1 (one FAH SMP client). If so, and system has 2 CPUs or 2 cores, then utility assigns two processes for each core in some magic way: first core gets process with maximal memory usage, and process with minimal memory usage. Other two processes are assigned to second core.

For quad core processors, first SMP client is assigned to 1st and 3rd core, 2nd SMP client assigned to 2nd and 4th core. Assigning affinities on QUAD if you run one SMP client does not make sense.
 
relic said:
I hear you Killer, but it does zero for me. Won't even set affinity.
Two different systems, Vista 64 and Vista x86. Nothing.

I gotta call shens.
Click to expand...

Not that I doubt your results Relic, but it works like a champ for me. Maybe your systems are so super optimized already that nothing can make them go faster :p
 
SpoogeMonkey said:
Not that I doubt your results Relic, but it works like a champ for me. Maybe your systems are so super optimized already that nothing can make them go faster :p
Click to expand...

Nah, he's just old and crotchety and can't figure out the newfangled technologies. Then again, Bill figured out the new technologies and he's even older and more crotchety. :D:p

 
My work computer with the affinity changer now lost the 100 ppd gain so it's not always as noticeable. It also depend on workunits.
 
I'll try this on the wife's rig tonight. Its a C2D so it should help it out some.

 
Xilikon said:
My work computer with the affinity changer now lost the 100 ppd gain so it's not always as noticeable. It also depend on workunits.
Click to expand...

Hey Xilikon:
I don't know if this is a "thread crap" or not :eek: , but do you think the time discrepencies are effected by the amount of L2 cache your cpu has or the type of cpu you have or the mobo chipset you have or your OC' or just the type of work unit? The reason I'm asking is I'd like to find the "optimum" settings and configaration for this SMP infinity deal, so far, so good. :D

From my "studious" reading on this forum the SMP infinity won't work on dual quad core cpus, etc Does anybody know, if any, the effect ithe program has on an AMD quad or Athlon64 x2? (does it even work on AMD cpus?)

Like some WU's worked faster with 4 MB L2 vs 2 MB L2 and some type of WU's worked faster with AMD cpus and some with Intel cpus. I wonder if there is a "sweet" spot for this infinity thing?? So far the time investment seems well worth it. :confused:

 
It is possible that the L2 cache also play a role. The work computers all have a E6300 with 2 Mb of cache. I dunno what is the specs of the new ones coming very soon but I bet they are 4 Mb (at least a E6420 hopefully).

 
I am running Vitsa on a T7200 laptop running a single smp client. And i noticed a difference using the AC proggie. My avg frames used to be about 22mins, currently i am watching my times for the last 1.5 weeks at around 20mins doing the same proteins.
 
I already posted this in the other Affinity Changer thread, but it works for me. Case in point, my E4300 saw a boost from 950 to around 1000-1050 PPD. Not a huge gain, but it is noticeable.

 
jws2346 said:
From my "studious" reading on this forum the SMP infinity won't work on dual quad core cpus, etc Does anybody know, if any, the effect ithe program has on an AMD quad or Athlon64 x2? (does it even work on AMD cpus?)

Like some WU's worked faster with 4 MB L2 vs 2 MB L2 and some type of WU's worked faster with AMD cpus and some with Intel cpus. I wonder if there is a "sweet" spot for this infinity thing?? So far the time investment seems well worth it. :confused:
Click to expand...
I read on several other forums that the Affinity Changer won't work on AMD rigs. Apparently, it makes things worse and lowers PPD. I haven't tried it yet, but I have three AMD systems including a dual dual-core (4 cores) with two SMP clients running and I'll report what I find.
 
SpoogeMonkey said:
Not that I doubt your results Relic, but it works like a champ for me. Maybe your systems are so super optimized already that nothing can make them go faster :p
Click to expand...

Naw, I don't think I could optimize Vista that well.

After a few hours of struggling with it I finally got it to set SMP client affinity on my Vista x86 system. This was my first problem, it would run as a service but it wouldn't do anything with the 2 SMP clients.

Had to leave for work so I can't watch it today, but afer running overnight it is still is within a few seconds per step of the same system without AC, so it's not improving the times but it is setting affinity. <shrug> I'll try it on the 64-bit machine tonight when I get home, The 64 bit Vista had one of the clients crash yesterday so AC wouldn't work.

We'll see if AC has any effect on that system, if not I'll just delete it and run the two SMP clients without it.
 
i know i know :(
but my ambient temp is around 30-33 c so with out an aftermarket hsf that's gonna stay stock.
working on getting a thermalright ultra 120


 
BillR said:
Please tell me that “stock Q6600” is a misprint or typing error. That would be like an act of heresy or sacrilege and could carry severe penalties.;):D
Click to expand...

The only time a Q6600 should be at stock in on the shelf.


 
No luck on the Vista x86.
Affinity is set but no change in step times.

Maybe if I change to XP and drop my Q6600 to "stock" ? :p

I'll setup the Vista 64 system after supper and see if there's any improvement there.
 
AC is working great on my system. My Q6600@2.88GHz was doing 1% per ~10 min on 64 bit vista (4Gb ram) with 1 SMP client. I loaded on AC and the 2nd SMP client and now each client runs at ~ 15 min per %. Twice the points with only a 50% time increase! It is running project 2653 now. It is even doing better on Winxp with my E6700@3.0GHz. When running the linux client (in Ubuntu 7.10) I was only getting about 1300-1400ppd now I am getting 2000ppd in win xp with AC.

You better get those Q6600s going Kendrak......I now have 4 SMP clients going strong now. FAH mon says it is pumping out about 7000 ppd....I don't know if I believe that, but it sounds like a lot. I still have another mobo that is just sitting on the shelf....so lonely ....just wanting one of those bargain Q6600s.

 
HighYield said:
AC is working great on my system. My Q6600@2.88GHz was doing 1% per ~10 min on 64 bit vista (4Gb ram) with 1 SMP client. I loaded on AC and the 2nd SMP client and now each client runs at ~ 15 min per %. Twice the points with only a 50% time increase! It is running project 2653 now. It is even doing better on Winxp with my E6700@3.0GHz. When running the linux client (in Ubuntu 7.10) I was only getting about 1300-1400ppd now I am getting 2000ppd in win xp with AC.

You better get those Q6600s going Kendrak......I now have 4 SMP clients going strong now. FAH mon says it is pumping out about 7000 ppd....I don't know if I believe that, but it sounds like a lot. I still have another mobo that is just sitting on the shelf....so lonely ....just wanting one of those bargain Q6600s.
Click to expand...

Yeah, but I seem to get the same performance with or without AC. I'm trying to figure out why people use it, apparently some see a boost in performance with it. I get nada.
 
With my quad, when i first started the other smp client 1% would take about 12 mins to complete, with both smps at first it took 20 mins then it went down to 17 mins per 1% on both clients. So now it takes wu 30 hours to finish instead of 20 hours but i get two wus finished.

 
HighYield said:
AC is working great on my system. My Q6600@2.88GHz was doing 1% per ~10 min on 64 bit vista (4Gb ram) with 1 SMP client. I loaded on AC and the 2nd SMP client and now each client runs at ~ 15 min per %. Twice the points with only a 50% time increase! It is running project 2653 now. It is even doing better on Winxp with my E6700@3.0GHz. When running the linux client (in Ubuntu 7.10) I was only getting about 1300-1400ppd now I am getting 2000ppd in win xp with AC.
Click to expand...
Not all of your increase in PPD is attributed to AC. Running two clients vs one client without AC is also more efficient. AC optimizes efficiency but is not solely responsible for your production increase.
 
APOLLO said:
Not all of your increase in PPD is attributed to AC. Running two clients vs one client without AC is also more efficient. AC optimizes efficiency but is not solely responsible for your production increase.
Click to expand...

What about the linux client on Ubuntu 7.10 running on the E6700@3.0GHz (2Gb ram)? It was only pulling about 1300-1400 ppd. I switched to WinXp pro with AC and now it is cranking 2000 ppd. I thought the the linux client was supposed to be more efficient. Is there anything special you are supposed to do on the linux client to get max ppd?


 
HighYield said:
What about the linux client on Ubuntu 7.10 running on the E6700@3.0GHz (2Gb ram)? It was only pulling about 1300-1400 ppd. I switched to WinXp pro with AC and now it is cranking 2000 ppd. I thought the the linux client was supposed to be more efficient. Is there anything special you are supposed to do on the linux client to get max ppd?


Click to expand...

My headless E4500 @3Ghz is pulling 2100 ppd with the F@H CD. Don't know what went on with yours.
 
HighYield said:
What about the linux client on Ubuntu 7.10 running on the E6700@3.0GHz (2Gb ram)? It was only pulling about 1300-1400 ppd. I switched to WinXp pro with AC and now it is cranking 2000 ppd. I thought the the linux client was supposed to be more efficient. Is there anything special you are supposed to do on the linux client to get max ppd?
Click to expand...
I know next to nothing about linux. I was only referring to quads and Windows clients. Try running two clients without AC and one client alone. You'll still see a production increase with 2 clients, and running AC will improve efficiency but it doesn't provide all of the increase in PPD...
 
HighYield said:
What about the linux client on Ubuntu 7.10 running on the E6700@3.0GHz (2Gb ram)? It was only pulling about 1300-1400 ppd. I switched to WinXp pro with AC and now it is cranking 2000 ppd. I thought the the linux client was supposed to be more efficient. Is there anything special you are supposed to do on the linux client to get max ppd?


Click to expand...

Some work units give better PPD than others. It sounds like the E6700 had one of the crappy ones which would explain the crappy PPD. From my experience that would probably be about the correct PPD for it on one of those units.

 
You must log in or register to reply here.
Back
Top