SMP Affinity Changer

[FunkZ]

I am always behind the times it seems, but better late than never right? This little gem of a program has been available for months however I only recently became aware of its existence and even then I was convinced by all the information available at the time that it would be of very little use on AMD processors,... dual-core's or quads running only 1 SMP instance especially.

The utility works by dynamically assigning each of the SMP Client's processes to a specific core based on its memory and CPU utilization, in order to ensure the client is working at maximum efficiency.

Testing on my 6400+ X2 rig is showing a consistent improvement in times. Prior to installation a Project 2665 (1920p) was running at an average 26:15/step, and is now completing at an average 23:30/step. That's a 2:45 decrease!

I have seen claims of 40-50% improvement in production using the Changer and 2 SMP instances on a quad core processor. There are some drawbacks as I understand from running 2 instances, and there is very little information regarding how it performs with AMD Phenom's. Following this trial with my X2 I plan to setup dual SMP's on my X4 to test with.

The SMP Affinity Changer can be downloaded from distributed.org (http://distributed.org.ua/forum/index.php?showtopic=1149)

[FunkZ]

It seems if you use the MPI version of the SMP client and run two instances it can create problems. I confirmed this on my X4, when one service was stopped it would hang the other one and the WU would EUE. However the Deino client appears to work just fine with two instances running, I am able to reboot and come right back to where they left off.

I am still monitoring the performance but it does appear that more work is being done, point production seems to be up. I do not have any of the old log files to compare frame times (the MPI client work is not compatible with the Deino client, so I had removed both folders prior to installing) however Fahmon is reporting a significant jump in PPD. Once it completes an entire WU I will post any gains.

[FunkZ]

http://home.comcast.net/~funkz/9848ppd.jpg
http://home.comcast.net/~funkz/10714ppd.jpg

Well, there you have it. The Affinity Changer increased point production by around 200 on the X2 and roughly 700 on the X4. That's a 23% and 41% jump respectively. Even if you only have a dual core processor, IMO the gain is worth downloading the program and installing it, the setup is very easy. Running two instances of the SMP on a quad core is a little more involved, but the greatest gains come from doing this, and the Affinity Changer helps to maximize the output of all four cores.

[slugbug]

A safer alternative to running two Windows SMP clients would be to install VMWare and run either one or two Linux virtual machines with Linux SMP clients.

[FunkZ]

VMWare runs around $200 though, doesn't it? Plus I would think you'd have extra overhead (memory + cpu) in the virtual OS footprint. Is Microsoft still giving away Virtual Server 2005? Pretty sure they have a Linux add-on pack for that.

The Deino client has been working flawlessly for me, running both SMP clients as a service, absolutely no EUE's since the update. Is this just abnormal or is there another reason it is safer to use Linux?

[slugbug]

VMWare runs around $200 though, doesn't it? Plus I would think you'd have extra overhead (memory + cpu) in the virtual OS footprint. Is Microsoft still giving away Virtual Server 2005? Pretty sure they have a Linux add-on pack for that.

The Deino client has been working flawlessly for me, running both SMP clients as a service, absolutely no EUE's since the update. Is this just abnormal or is there another reason it is safer to use Linux?

VMWare Server and VMWare Player are free to use. Scroll down to LumberJack's post #19 and #20 for an excellent guide on "Running notfred's in VMWare Player, with a USB flash drive, that saves and restores." A little over half way down the page.

http://techreport.com/forums/viewtopic.php?f=9&t=59595