Sluggish SMP Performance

[Tytalus]

At work I have a machine with the following specs:

Pentium D 920 (Presler core, 65nm, 2.80GHz) which supports the EM64T instruction set. It has 2GB of RAM in it and that's about all I know (it's a Dell, not that the info is useful.) I'm incapable of dismantling or in any way modifying the hardware of the PC.

I loaded VMware and got everything up and running with a 64-bit install of Red Hat and the SMP client. I let the SMP client run overnight but the performance is less than stellar! I gave the virtual machine 1024MB of RAM so it would have a lot of room to work with and the performance was on the magnitude of 1.25 hours per percent.

A quick snippet:

[21:59:22] Completed 0 out of 500000 steps (0 percent)
[22:36:38] Writing local files
[22:36:38] Completed 5000 out of 500000 steps (1 percent)
[23:48:18] Writing local files
[23:48:19] Completed 10000 out of 50000 steps (2 percent)
[01:35:19] Writing local files
[01:35:20] Completed 15000 out of 500000 steps (3 percent)
[03:24:48] Writing local files
[03:24:48] Completed 20000 out of 500000 steps (4 percent)
[05:15:05] Writing local files
[05:15:05] Completed 25000 out of 500000 steps (5 percent)
.....
.....
.....

The performance actually seems to be getting WORSE over time. Any suggestions on improving the performance? I wouldn't think that it would take 8-9 days to complete a work unit on a machine like this!

 

[zim01]

This is windows based then?

 

[Tytalus]

Zim:

I am running a 64-bit Red Hat server in VMware on a machine using Windows XP SP2. Thanks for the question--I didn't realize that I hadn't clarified that!

 

[zim01]

PM:unhappy_mage

 

[marty9876]

If it's a WU 2605 these are known to run like shit on P4D's, check the offical forums.

Post WU project number.

 

[Tytalus]

I didn't even think to check the work unit....

Project: 2604 (Run 0, Clone 201, Gen 10)

....think maybe you meant 2604? Either way I'm gonna' nuke the work unit and get another one sent out.

Thanks for the heads-up!

 

[marty9876]

Copy/pasted from a thead on FCF:

2651 has the same points as the 2604/2605 and the same deadline, yet it folds 3 minutes slower on a frame.

2604/2605 on Ubuntu/VMware: 12:50
2651 running on Win SMP: 15:50

All folded on the same computer (C2D E6600@3.0)

I tend not to follow PPD's across many WU sets sio I'mm bow out now. Unless your going to miss the final deadline you generally should not dump WU's.

 

[Tormond]

On my C2D 6400@ 3.2 I get about 13.5 min per frame in VMWare using mages install. At those time rates you aren't going to make deadlines I dont believe but I could be wrong..I am the suck at math

Edit: This is on 2604 WU

 

[[BRO]Alaskan]

1hr 11min 40 sec will have that WU complete in 4.97 days. thats .97days to long

 

[Tigerbiten]

I've seen the recomended speed for a P4D to make the deadlines is ~2.5 Ghz.
So add the 20-25% speed reduction due to the VMWare and your then looking at just over 3 Ghz to make the deadlines.

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

 

[Tytalus]

I've seen the recomended speed for a P4D to make the deadlines is ~2.5 Ghz.
So add the 20-25% speed reduction due to the VMWare and your then looking at just over 3 Ghz to make the deadlines.

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

*SIGH* Yeah, I thought as much. Today it's even worse.... looking at over 2 hours per frame to make it happen.

It appears I'll have to bow out of the SMP market until some of these WU's get doable on my hardware using VM. I don't have the option of switching over to something else so that's that. Looks like Adrenaline will overtake me in about 1.5 weeks then at the rate he's chugging. Damn his work for giving him a laptop he can fold on! *SHAKE FIST*

EDIT: Oh, I tried dumping the WU and I kept getting the same one...? I wiped the work directory and nuked the FahCore_a1.exe file on the Linux box, but it kept downloading the same core (which sorta' makes sense) then it kept working on the same gen/clone/etc.? Is there something I'm missing?

 

[Tigerbiten]

If the client think you have had an EUE then it will download the same protien twice more as it needs to check its the correct result.
Its after three attempts to finish the protien you get a new one.

Its happened to me once so far and I got no points for 24 hours folding as the client was deleting the work unit before asking for a new one and not sending the part work in.

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

 

[Tytalus]

Wow.... that's really a pain in the rear! I deleted that work directory several times in a row, is there a better solution to reliably receiving a "new" work unit? I'd really like to skip this one and attempt one that I can do within the 4-day period (well, it'd be nice to get one I could get done in like 1-2 days.... but I'm not going to be THAT choosey =] )

Just want to help where I can! (Well, where my work PC can!)

Oh, FYI--got my old home PC working. Sadly it's only an AMD Athlon 64 3800+ (single core) and for some reason the SMP clients really don't like it. I'll have to grab a non-SMP 64-bit client and run it in VMware on there (can't nuke it, still occasionally used for Windows-based stuff) and see what-all I can get performance-wise. For now it's using the standard F@H console-based client as a service. It's "monitor" is my 50" Sony TV. =)

 

[Mohonri]

Oh, FYI--got my old home PC working. Sadly it's only an AMD Athlon 64 3800+ (single core) and for some reason the SMP clients really don't like it. I'll have to grab a non-SMP 64-bit client and run it in VMware on there (can't nuke it, still occasionally used for Windows-based stuff) and see what-all I can get performance-wise. For now it's using the standard F@H console-based client as a service. It's "monitor" is my 50" Sony TV. =)

The SMP client only works on SMP machines, which yours is not. And because of that, there's no need to run VMWare on your machine. Just run the standard console service--that's the best way for your CPU.