Is the 96 protein fold slower for you guys too?

[Davdog]

Here's a newbie sort of question. I was crunching about 3.3 structures per second with the 108, the 96 is about 2.6 s/s. Should I assume that there is something different with this protein that makes it slower to run?

Dave

[bwkaz]

If the protein structure makes atoms "run into" each other more often, then the algorithm has to backtrack more often, and the protein will run slower. That's at least one reason why a shorter protein can be slower...

[Davdog]

Its been a while since I did organic chem, but i do remember the concept of steric hindrance, and figured that could have something to do with it.

[Darkness Productions]

An XP2100 running at 1805:

Old protien:


------------------------------------------------------------
Distributed Folding Windows dfGUI v1.8 Benchmark

Sample Size : 193800 structures over 46738 seconds.
Protein Size: 108AA

Structures Per Second: 4.15
Structures Per Minute: 248.8
Structures Per Hour : 14927
Structures Per Day : 358259

OS : Windows 2000 MHz: 1805
CPU: AMD Athlon(tm) XP 2200+
Client Switches: -df -qt -rt -g 15 -s 10000
------------------------------------------------------------


New protien:


------------------------------------------------------------
Distributed Folding Windows dfGUI v1.8 Benchmark

Sample Size : 1995 structures over 750 seconds.
Protein Size: 96AA

Structures Per Second: 2.66
Structures Per Minute: 159.6
Structures Per Hour : 9576
Structures Per Day : 229824

OS : Windows 2000 MHz: 1805
CPU: AMD Athlon(tm) XP 2200+
Client Switches: -df -qt -rt -g 15 -s 10000
------------------------------------------------------------


Glen

[Davdog]

old =
Distributed Folding Windows dfGUI v2.1 Benchmark

Sample Size : 247815 structures over 73120 seconds.
Protein Size: 108AA

Structures Per Second: 3.39
Structures Per Minute: 203.3
Structures Per Hour : 12201
Structures Per Day : 292823

OS : Windows XP MHz: 1536
CPU: AMD Athlon(TM) XP 1800+
Client Switches: -df -qt -rt -g 15 -p 10

new
Distributed Folding Windows dfGUI v2.1 Benchmark

Sample Size : 70605 structures over 29855 seconds.
Protein Size: 96AA

Structures Per Second: 2.36
Structures Per Minute: 141.9
Structures Per Hour : 8514
Structures Per Day : 204330

OS : Windows XP MHz: 1536
CPU: AMD Athlon(TM) XP 1800+
Client Switches: -df -qt -rt -g 100 -p 10

[IronBits]

How many AMD GHz does it take to crank 1 million structures per day? Intel P3? Intel P4?
I only want the AMD, but figured what the heck, if someone knows. ;)
That would be 1,000,000/24 = 41667 (rounded up) per hour.

[Darkness Productions]

From one of my linux boxen -

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

Sample Size: 598960 structures over 233069 seconds.

Structures Per Second: 2.57
Structures Per Minute: 154.19
Structures Per Hour: 9251.58
Structures Per Day: 222037.92

Linux OS (grant) - Running Kernel Version 2.4.20
AMD Athlon(tm) XP @ 1636mhz (256 KB cache)
------------------------------------------------------------


So assume just sub-5 of those, or about 8GHz, give or take a bit.

[IronBits]

Thanks DP!

[Angus]

P4 2.0 GHz, 512M of DDR ram
------------------------------------------------------------
Distributed Folding Windows dfGUI v1.9 Benchmark

Sample Size : 775 structures over 352 seconds.
Protein Size: 96AA

Structures Per Second: 2.20
Structures Per Minute: 132.1
Structures Per Hour : 7926
Structures Per Day : 190227

OS : Windows 2000 MHz: 1993
CPU: Intel(R) Pentium(R) 4 CPU 2.00GHz
Client Switches: -df -qt -rt -s 1000
------------------------------------------------------------

so roughly 5 of these (10 GHz) for your million-a-day mark



Single instance of DF on dual P4 Xeon 2.8 GHz, 2GB SDRAM
------------------------------------------------------------
Distributed Folding Windows dfGUI v2.1 Benchmark

Sample Size : 635 structures over 283 seconds.
Protein Size: 96AA

Structures Per Second: 2.24
Structures Per Minute: 134.6
Structures Per Hour : 8078
Structures Per Day : 193866

OS : Windows 2000 MHz: 2790
CPU: Intel(R) Xeon(TM) CPU 2.80GHz
------------------------------------------------------------

As you can see there is a significant performance difference that can only be attributed to memory type. This is born out by the results from my P4 2.6 GHz SDRAM home machine (benchmark later), which is slower than the 2.0 GHz P4 with DDRAM. Same major brand of PC, same OS.


However! running 4 instances of DF on the P4 Xeon duallies using HyperThreading results in quite a bit more output than just running 2 instances. The duallie then puts out about 640K per day, from 5.6 GHz. That would only require about 8.75 GHz of HT Xeon to achieve the 1 million a day mark.

HTH

Angus

[Paratima]

Originally posted by IronBits
Thanks DP! Yes. Thanks, indeed! :D

[TheOtherPhil]

Originally posted by Angus
However! running 4 instances of DF on the P4 Xeon duallies using HyperThreading results in quite a bit more output than just running 2 instances. The duallie then puts out about 640K per day, from 5.6 GHz. That would only require about 8.75 GHz of HT Xeon to achieve the 1 million a day mark.

HTH

Angus


I am assuming that this is dual 2.8 Xeons? Very interesting. I have been toying with the idea of a dual Xeon box (I run 5 dual AMD systems) but was unsure of the performance to expect. I am currently getting ~550K from dual XP2400's at 2250MHz, so the Xeons don't look too bad a purchase.

Thanks.

[Angus]

I am assuming that this is dual 2.8 Xeons?

Yes. It would be a real screamer with DDR memory.

HT seems to work well with DF, but it pukes on RC5-72.

The DNET RC5-72 client sees 4 procs, and starts 4 crunchers, but the combined rate is very slow - about one CPU worth. Manually configuring it to run 2 crunchers gets the best result.

[Alpha_7]

Oh to have dual processes...


One day, one day :D

[Davdog]

When I figured out the costs to set up a dualie, it seemed just as effective to get single units, of course, the cool factor is significantly lower!

Dave

[TheOtherPhil]

Originally posted by Davdog
When I figured out the costs to set up a dualie, it seemed just as effective to get single units, of course, the cool factor is significantly lower!

Dave


I worked it out differently....1 PSU, 1 GFX card, 1 HD, 1 CDROM etc, etc. The fact that I can get more CPU power in less boxes also means my wife doesn't go mental. She's bad enough with 5 machines spread about the house....10 singles and we'd be divorced :jester:

[TheOtherPhil]

Originally posted by Darkness Productions
An XP2100 running at 1805:

New protien:


------------------------------------------------------------
Distributed Folding Windows dfGUI v1.8 Benchmark

Sample Size : 1995 structures over 750 seconds.
Protein Size: 96AA

Structures Per Second: 2.66
Structures Per Minute: 159.6
Structures Per Hour : 9576
Structures Per Day : 229824

OS : Windows 2000 MHz: 1805
CPU: AMD Athlon(tm) XP 2200+
Client Switches: -df -qt -rt -g 15 -s 10000
------------------------------------------------------------


Glen

Looks like that box is underperforming a little Glen. I get the following on dual XP1800's running Gentoo:


CPU0
Sample Size: 115100 structures over 44236 seconds.

Structures Per Second: 2.60
Structures Per Minute: 156.12
Structures Per Hour: 9367.03
Structures Per Day: 224808.72


CPU1
Sample Size: 116300 structures over 44718 seconds.

Structures Per Second: 2.60
Structures Per Minute: 156.04
Structures Per Hour: 9362.67
Structures Per Day: 224704.08

[Darkness Productions]

That box is windows, which has been known to underperform compared to Linux. One of these days, I may put linux on it to see how well it does...