New Windows SMP client release (Beta 5.92): Read All Directions

A

aldamon

Supreme [H]ardness
Joined
May 24, 2000
Messages
6,671
http://foldingforum.org/viewtopic.php?f=8&t=1783

We are excited to release a new Windows SMP client to beta. This is an *early* beta release--please expect some rough edges.
Read below for details:

*how to get it*
go to the download page: http://folding.stanford.edu/English/Download (substitute your language as appropriate) and look for the 5.92 client

***how to install it***
The installation process is a little complicated. PLEASE READ BELOW.
1. The initial release is a *replacement* for your current 5.91 SMP client. We will have a fresh install version soon.
2. Make a backup of a working FAH Windows/SMP installation. Install into your current SMP working directory
3. Unzip the 5.92 file in your test directory
4. Open a command prompt window (run as administrator if Vista)
5. run install.bat. This will do several things
.. a) install DeinoMPI. You don't need to do anything special here
.. b) create a "credential store" that DeinoMPI will use to authenticate the mpi processes. Important: do not use a passphrase here (just hit return for no passphrase and say yes, you're sure). I recommend encrypting the data and using the registry for storage.
.. c) add your user information to the encrypted credential store. This will prompt you for your username and password.
.. d) run the standard mpiexec -np 2 foo test application to make sure things work. On some WinXP systems, you may see a defective Windows Firewall dialogue box come up. The firewall may also ask you to unblock applications (this is a good idea).
Notes from the forum mods on the above process:
5.d) The Windows Firewall dialog box will pop up twice. The first time, you will be able to see and click Unblock for DeinoMPI. The second time, the pop up window is hosed. However, since the unblock button has a hot key, you can press alt+u to unblock the next item and make it go away. I checked my firewall exceptions list after unblocking both. The second pop up is for fahcore1.exe.
5.d.2) I found a different bypass. After having gotten the hosed windows many times in WinXP I went to Start + ControlPanel + SecurityCenter + WindowsFirewall. On the Exceptions tab I scrolled through the list of names, adding a checkmark to the appropriate boxes.
6. Start fah.exe as normal.
7. Please report problems to this forum.

*what's different in 5.92*
With version 5.92, we have switched to a different MPI implementation (DeinoMPI instead of MPICH). In our testing thus far, this new implementation is *much* more stable for FAH. Performance is roughly comparable, with perhaps a slight speed increase. But stability and reliability is what we're going for here.
DeinoMPI is backwards-compatible with MPICH, so your old core binaries will work. However, most of the benefits come with a new core binary. One is included in the install package; the client should also auto-download Deino-specific core binaries. Similarly, the mpiexec binary for DeinoMPI differs from MPICH.
Look for new Deino-specific features coming in the future.

*other notes*
5.92 does not currently support 64-bit Windows. This is something we will work on in the future.
As with 5.91, we do not yet officially support running 5.92 as a service. It is possible to do this, but it requires setting several dependencies, etc. Other FAH donors may have helpful hints here.

Thanks for testing!
Click to expand...


 
Geez, not even mentioned on folding.typepad.com yet.
Something new to play with. Hopefully this ends the need for the loopback adapter install.
 
I dunno. I just stumbled onto it at the forum.

I just installed it without problems on my work computer. It started from the last checkpoint and didn't trash the work unit. Yay!

 
I'll try that on my E2160 box tonight (win XP 32 bits since the E2160 doesn't support 64 bits virtualization). If the performance is better and is more reliable, I'll upgrade all my SMP boxes at work (they are all XP-32).

 
aldamon said:
I dunno. I just stumbled onto it at the forum.

I just installed it without problems on my work computer. It started from the last checkpoint and didn't trash the work unit. Yay!

Click to expand...

Try pulling the network cable/killing the wifi and seeing if it stops the Fahcores.
 
Looks to be much more involved with the installation process. Other than a slightly increased stability, there's no major incentive to reinstall all my SMP clients to this new beta. I might try it on completely new builds or OS installations, but I'm leaving my current clients alone. I give credit to Stanford for making any effort to improving the client, though.

 
Um, its no more involved than the normal 5.91 install really. Only issue if you're running Windows Firewall really.
 
DriveEuro said:
Try pulling the network cable/killing the wifi and seeing if it stops the Fahcores.
Click to expand...

You first. :p

7im seems to think that little problem is fixed though.

And yes, this MPI stack does seem to be more tolerant of network changes. If you don't believe me, just unplug your network drop and watch what happens. Don't freak at first, just continue watching. :twisted:
Click to expand...



 
aldamon said:
I dunno. I just stumbled onto it at the forum.

I just installed it without problems on my work computer. It started from the last checkpoint and didn't trash the work unit. Yay!
Click to expand...

Ditto here - going to do the dual xeon box in a few - I'll post if I see any problems.



 
The sample size is small so far, but early results show a large drop in PPD with this client (Dell E6700 @ stock, 2653 unit).

I just rebooted and hopefully that fixes it. I'm going to let it run overnight here before I upgrade at home.



 
aldamon said:
The sample size is small so far, but early results show a large drop in PPD with this client (Dell E6700 @ stock, 2653 unit).

I just rebooted and hopefully that fixes it. I'm going to let it run overnight here before I upgrade at home.
Click to expand...
Aldamon, if you see a significant drop in production, don't let it run all night. It only requires 3-4 frames before we get an accurate estimation of actual PPD, providing nothing else is running at the same time. I have been conducting tests on multiple system configurations, with all kinds of combinations of hardware and BIOS settings for over a week now in order to get the maximum out of all my S771 hardware. I let the clients work on only a few frames for each combination and then move on to the next permutation. If I let it run longer, I don't see any difference in output.
 
I'm willing to run it overnight in order to provide constructive criticism. If I dump it after an hour they probably won't take me seriously, you know? I am going to report the drop in PPD as soon as the next frame finishes.

 
This isn't 64bit compatible either, so that counts out my main boxen. I think I'll sit on this until someone comes back with some solid testing results.

 
Axdrenalin said:
This isn't 64bit compatible either, so that counts out my main boxen. I think I'll sit on this until someone comes back with some solid testing results.

Click to expand...

Yeh.... I'm still running and haven't had to many problems, I'll wait.

 
If I see some indications of improvement in production, I'll try it on one of my dual socket-A machines since they could use all the boost they can get. Until then, I'm sitting tight.
 
I'll be giving it a try too - i had nothing but problems with the older windows smp client. I'm running two VMs on this new quad, but I don't like the overall loss of responsiveness of my system - one VM and one SMP means that I can do other things without shutting down one of the VMs.

 
Hopefully it won't take them too long to get some 64 bit support. It would be great if they got this software a bit more robust. Anybody notice any changes in the other verisions of the client software for other platforms?

 
Won't run for me, keeps saying unexpected termination of Folding@home core, after about 3 tries, it downloads a new core, and repeats. Went back to 5.91.
 
aldamon, I remember someone saying the Pande group is trying the new client to see if the new MPI works better than the current MPICH implementation. with your testing, I guess i'll leave my WinSMP boxes alons and just add the loopback adapter...

 
Yep - think I'll hold off on upgrading the dual quad box till I see more results on this. I did install it on the E6550 dual core box and it seems to be running fine. I did not bother to see what my prior frame times were though. Right now it's crunching a 2653 work unit at about 25 minutes per frame. No idea if thats good or not. That system is a bone stock Dell so it's not overclocked at all.

 
To add to my comment, I'm already contributing on the new A2 Linux core and 2619 WU in the private beta forum (I'm a beta tester). I already provided the comments from the Pande group about that so we'll see if they will adjust the WU points with our feedback along with the upload parameters.

 
Logan321 said:
Won't run for me, keeps saying unexpected termination of Folding@home core, after about 3 tries, it downloads a new core, and repeats. Went back to 5.91.
Click to expand...


I'm getting the same error. Went back thru the installation instructions - afaik I did everything correctly.

oh well - i'll try again if they come out with a new beta, but otherwise its back to two VMs.
 
MixManSC said:
Yep - think I'll hold off on upgrading the dual quad box till I see more results on this. I did install it on the E6550 dual core box and it seems to be running fine. I did not bother to see what my prior frame times were though. Right now it's crunching a 2653 work unit at about 25 minutes per frame. No idea if thats good or not. That system is a bone stock Dell so it's not overclocked at all.

Click to expand...

Is that running SMP Affinity changer? I'm running a 2653 WU on a Dell with a E6400 and I'm getting 18-19 minutes per frame on the old client

 
No..... should I be running affinity changer? It's only a dual core box running one smp client.


Crazy Larry said:
Is that running SMP Affinity changer? I'm running a 2653 WU on a Dell with a E6400 and I'm getting 18-19 minutes per frame on the old client

Click to expand...
 
My E6600 @ 3.6ghz averages about 10m 50sec per frame with affinity changer.

 
Damn - I've been missing out. Just installed AC on the Dell and will also install this in the C2D at home later. I'll post in a few hours what gains I got on the Dell.
 
aldamon said:
Could you look at your log and compare the before/after frame times please?


Click to expand...


It actually appears to be a few seconds faster looking at the log for the 2653 work unit it was in progress on when I changed it to the new client. Was getting about 24.5 minutes per frame on 5.91 beta. After installing 5.92 beta is was averaging 23.5 minutes per frame. Then it finished and downloaded a new 2653 wu and it's averaging 25 minutes per frame.
 
I don't understand how your E6550 can be so slow with either client. I get ~15:40 per frame with 5.91 and 2653. Here are my comparison numbers:

E6700 @ stock (2.66 GHz) on 2653

Overnight times with 5.91
[04:58:34] Writing local files
[04:58:34] Completed 110000 out of 500000 steps (22 percent)
[05:13:14] Writing local files
[05:13:14] Completed 115000 out of 500000 steps (23 percent)
[05:27:58] Writing local files
[05:27:59] Completed 120000 out of 500000 steps (24 percent)
[05:42:38] Writing local files
[05:42:38] Completed 125000 out of 500000 steps (25 percent)

Overnight times with 5.92
[23:20:51] Writing local files
[23:20:51] Completed 460000 out of 500000 steps (92 percent)
[23:37:03] Writing local files
[23:37:03] Completed 465000 out of 500000 steps (93 percent)
[23:53:08] Writing local files
[23:53:08] Completed 470000 out of 500000 steps (94 percent)
[00:09:16] Writing local files
[00:09:16] Completed 475000 out of 500000 steps (95 percent)

More Overnight 5.92
[10:15:40] Writing local files
[10:15:41] Extra SSE boost OK.
[10:31:55] es
[10:31:55] Completed 5000 out of 500000 steps (1 percent)
[10:48:04] Writing local files
[10:48:04] Completed 10000 out of 500000 steps (2 percent)
[11:04:16] Writing local files
[11:04:17] Completed 15000 out of 500000 steps (3 percent)
[11:20:25] Writing local files
[11:20:25] Completed 20000 out of 500000 steps (4 percent)
[11:36:37] Writing local files
[11:36:37] Completed 25000 out of 500000 steps (5 percent)
[11:52:46] Writing local files
[11:52:46] Completed 30000 out of 500000 steps (6 percent)

Switched back to 5.91
[12:21:51] Read checkpoint
[12:21:51] out of 500000 steps (6 percent)
[12:21:51] Extra SSE boost OK.
[12:21:51] d 30000 out of 500000 steps (6 percent)
[12:21:52] Extra SSE boost OK.
[12:37:40] Writing local files
[12:37:40] Completed 35000 out of 500000 steps (7 percent)
[12:53:08] Writing local files
[12:53:08] Completed 40000 out of 500000 steps (8 percent)
[13:08:48] Writing local files
[13:08:48] Completed 45000 out of 500000 steps (9 percent)
[13:24:24] Writing local files
[13:24:24] Completed 50000 out of 500000 steps (10 percent)
[13:40:01] Writing local files
[13:40:02] Completed 55000 out of 500000 steps (11 percent)

 
5.92 is good overall

I've lost about 200PPD from each SMP, but it is less prone to network hiccups. I would say the increased stability out weighs the minor loss of PPD. Considering you lose less than 5%, but gain stability. 1 outtage a week can easily run you more than that 5% decline. I'm going to give it longer to run on my 3.6ghz Quad and report back if my frames increase

Before:

10m 45s per frame

Now:

12m per frame to 12m20s per frame

Drop from 2350 PPD to 2150PPD

So its a judgement call.

One other side note the new deniompi (sp) takes up alot of memory its taking about 14mb per core x8 = 112mb of memory used by the new mpi. I don't recall the old mpi taking that much memory, do you? Infact it took about 2mb before so new client = 110mb diff in memory usage!

I was able to verify that the cores do keep going regardless of the network cable and if it is about to download or send it just keeps trying about every 9min. So very nice improvement.

 
I would rather use 5.91 + installing loopback adapter (or use Vista) than lose PPD with a inefficient beta.

 
Xilikon said:
I would rather use 5.91 + installing loopback adapter (or use Vista) than lose PPD with a inefficient beta.

Click to expand...

Vista has displayed the lost network client hang-up just as XP has for me at least.
 
Sunin said:
One other side note the new deniompi (sp) takes up alot of memory its taking about 14mb per core x8 = 112mb of memory used by the new mpi. I don't recall the old mpi taking that much memory, do you? Infact it took about 2mb before so new client = 110mb diff in memory usage!
Click to expand...
I never seen this process - must be in the new client. The mpiexec consumes about 4080k each and there are four of them running in my system with two clients. So, an approximate total of 15.8MB of memory used. That's a major jump in resource consumption with the new client if confirmed. I don't wish to install this on my systems with low memory. Stanford better improve it before the next expiration date if it's slated to be a replacement for the current client.

Xilikon said:
I would rather use 5.91 + installing loopback adapter (or use Vista) than lose PPD with a inefficient beta.
Click to expand...
I agree, besides the network cut off is rare for me and I use it to advantage when I want to shut down multiple clients quickly and risk free. I don't know of a better way to do this.

 
Sunin, not for me... I yanked the network cable off my main box with Vista without killing the client ;)

If you still stall the client with Vista when unplugging, something is badly configured ;)

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