performance on an alpha

[Kaliban]

Hi all.

Just wondering how well DF runs on an alpha machine. I could probably set it up on an idle ES40 (quad 667Mhz EV67 cpu's) at work, but I really have NFI when it comes to unix or vms so I'd have to go running off looking for help at setting it up.

Anyone here run it on an alpha?

[Paratima]

Supposed to be VERY fast, due to Alpha's architecture vs. the algorithm the client uses. However, the speed of running this client is almost wholly tied to Mhz, and you have to put out a lot of $$ to get Mhz with Alphas. Of course, if you've got one or two at work... toot sweet!

There are prior threads here (http://www.free-dc.org/forum/showthread.php3?s=&threadid=1064) and here (http://www.free-dc.org/forum/showthread.php3?s=&threadid=813).

I haven't done it, but xman has, maybe he'll weigh in with some wisdom. Good luck! :cool:

[vsemaska]

I run the client on Alphas and yes it is fast. The fact that you have a *quad* system means you generate a hell of a lot of structures. (You'll have to run 4 instances of the client, each one in its own directory).

You don't have to worry about VMS since the client doesn't run on it. You'll have to run in on (probably Tru64) UNIX. It's not that bad.

Vic

[MAD-ness]

I want to see performance numbers when you get it up and running.

Even at P3 performance levels (at those clock speeds) you would see pretty darn good results from 4 CPUs.

DF seems to have little performance impact when when running two instances on two cpus, but I don't know how a 4-way system would fare. My guess is that with suffiicient RAM you would still have good performance out of all 4.

[Kaliban]

Ahhh, cool!

Since I work in a classic DEC integration centre this is a test machine we use from time to time. Just hasn't been used in a long time since the alphas aren't exactly flying out the door :rolleyes: :(

[MAD-ness]

If you made it "Fly out the door" to my house, I would be much obliged. ;)

[xman]

Well, here is the particular instance of the speed of DF on alphaserver, EV6.8C, 1GHz with 8MB L2 Cache. However, it is found not better than my AMD MP 1.2GHz. But the structures per second is not always constant, it keeps changing, but it is around 1.2 to 1.4 on the alphaserver.

What I think is that, the final speed of the DF is still rely on how optimized is the program on that platform. Thus, alphaserver is not always faster than any other platform, even though it is faster in some cases, with lower clock speed.

This is running the 111AA, current protein, with -df -qt -rt -i f

------------------------------------------------------------
Distributed Folding Linux Benchmark Script V1.0

Sample Size: 720 structures over 542 seconds.

Structures Per Second: 1.33
Structures Per Minute: 79.70
Structures Per Hour: 4782.29
Structures Per Day: 114774.96

Linux OS - Running Kernel Version V5.1
@ mhz ( cache)
------------------------------------------------------------

------------------------------------------------------------
Distributed Folding Linux Benchmark Script V1.0

Sample Size: 3835 structures over 2740 seconds.

Structures Per Second: 1.40
Structures Per Minute: 83.98
Structures Per Hour: 5038.69
Structures Per Day: 120928.56

Linux OS - Running Kernel Version 2.4.19
AMD Athlon(tm) Processor @ 1194mhz (256 KB cache)
------------------------------------------------------------

[MAD-ness]

Thanks for the performance numbers xman.

BTW, nice production. :)