QMD's rock hard.

[Revv23]

QMD's rock hard.

Ok so i had a gromacs core that took 25 hours to complete, at it wsa 157 points, im working on a QMD now thats worth 450 points in about 12 hours total.


If your computer can handle QMD's, make sure you ask for them!!!

[Ugly n Grey]

Yeah it makes a huge difference. I am going to try a couple Dual Intel Processor rigs and see how they fare.

[Revv23]

Wow, seriously i think it ended up being 13 hours total...

and i got assigned another QMD core now...

according to FAHMON im putting out about 800pts per day just on this one CPU with QMD's.

Hehe im pumped... should have two more cpu's online by the end of the weekend too btw.. im surprised how many of my college buddies have said no, they trust me to fix thier computers, but for some reason the thought of this running in the background doesnt appeal to them... maybe i should set them up with graphical clients so they can watch it themselves lol?

[Ugly n Grey]

That's a good idea, or invite them to drop by XS and see what the score is themselves :)

Nicely done on the scoring btw very sweet indeed!

[Revv23]

heh not sure if ill show them XS, i might not have any friends after that, i let on to my friends that i know a thing or two about computers, not that i am a hardcore freak about them lol.

[angra]

I have been trying to get a solitary machine up & running on QMDs.

I keep getting a Gromacs core - FahCore_78.exe. That's not QMD is it?

I have the config option to accept large packets, running with -advmethods command line option, CPUs are Xeons.

Am I missing something?

This machine is thermally isolated, and is "busy hiking through the andes" student's workstation, so it shouldn't wig out at the extra load, so I'd really like this one to be getting good points for the team...

[Ugly n Grey]

Hmmm, sometimes if certain units were aborted on a box, you can have problems. You can experiment by entirely deleting the existing install and starting from scratch on that unit. But...

Before you do that, can you email the current logs of that unit to me at "uglyngrey hotmail" and I will give em a glance and see what's up?

It may be a simple solution.

[angra]

Hmmm, sometimes if certain units were aborted on a box, you can have problems. You can experiment by entirely deleting the existing install and starting from scratch on that unit. But...

Before you do that, can you email the current logs of that unit to me at "uglyngrey hotmail" and I will give em a glance and see what's up?

It may be a simple solution.

I'll post here in case the analysis is useful to others. If this is inappropriate for some reason (too spammy, personal info buried in there or whatever) please feel free to delete, Mr. Moderator :)


[13:24:26] - Ask before connecting: No
[13:24:26] - User name: Angra (Team 36362)
[13:24:26] - User ID not found locally
[13:24:26] + Requesting User ID from server
[13:24:27] - Machine ID: 1
[13:24:27]
[13:24:27] Work directory not found. Creating...
[13:24:27] Could not open work queue, generating new queue...
[13:24:27] + Benchmarking ...
[13:24:32] - Preparing to get new work unit...
[13:24:32] + Attempting to get work packet
[13:24:32] - Connecting to assignment server
[13:24:32] - Successful: assigned to (171.64.122.130).
[13:24:32] + News From Folding@Home: Welcome to Folding@Home
[13:24:33] Loaded queue successfully.
[13:24:36] - Deadline time not received.
[13:24:37] + Closed connections
[13:24:37]
[13:24:37] + Processing work unit
[13:24:37] Core required: FahCore_78.exe
[13:24:37] Core not found.
[13:24:37] - Core is not present or corrupted.
[13:24:37] - Attempting to download new core...
[13:24:37] + Downloading new core: FahCore_78.exe
[13:24:38] + 10240 bytes downloaded
....
[13:24:40] + 1134362 bytes downloaded
[13:24:40] Verifying core Core_78.fah...
[13:24:40] Signature is VALID
[13:24:40]
[13:24:40] Trying to unzip core FahCore_78.exe
[13:24:41] Decompressed FahCore_78.exe (3435392 bytes) successfully
[13:24:41] + Core successfully engaged
[13:24:46]
[13:24:46] + Processing work unit
[13:24:46] Core required: FahCore_78.exe
[13:24:46] Core found.
[13:24:46] Working on Unit 01 [October 21 13:24:46]
[13:24:46] + Working ...
[13:24:46]
[13:24:46] *------------------------------*
[13:24:46] Folding@Home Gromacs Core
[13:24:46] Version 1.86 (August 28, 2005)
[13:24:46]
[13:24:46] Preparing to commence simulation
[13:24:46] - Looking at optimizations...
[13:24:46] - Created dyn
[13:24:46] - Files status OK
[13:24:47] - Expanded 461276 -> 2782852 (decompressed 603.2 percent)
[13:24:47] - Starting from initial work packet
[13:24:47]
[13:24:47] Project: 759 (Run 6, Clone 3, Gen 0)
[13:24:47]
[13:24:47] Assembly optimizations on if available.
[13:24:47] Entering M.D.
[13:24:54] Protein: p759_vln3.8-0.2nopbc
[13:24:54]
[13:24:54] Writing local files
[13:24:54] Extra SSE boost OK.
[13:24:54] Writing local files
[13:24:54] Completed 0 out of 250000 steps (0%)

[Ugly n Grey]

It's all good I thought the log might be huge :)

OK something I missed was that's only your first packet. A new P4 type core will rarely get a QMD on the first try. The client stores sucesses and failure to match against future possibilities when left on it's own. As well do try and go for -forceasm and see what happens, this forces the client not only to use all available optimizations, it prevents it from examining/looking back at past failures (of which there are none on a new rig) and prevents the client from needing to establish a baseline.

As well, the computer needs to report back 512Meg of RAM or greater to get QMD's as far as I understand it.

so basically, after indicating large work units in the setup portion, start the client with the following flags
-svcstart(if it is the command line version) -local -advmethods -forceasm -verbosity 9

Hope this helps, your next packet should be a QMD
Edit: Oh what Xeons are these? you can run CPUID if need be to see what extensions the processor has (ie SSE2)

[angra]

no cpuid on this machine but /proc/cpuinfo:


processor : 0
vendor_id : GenuineIntel
cpu family : 15
model : 2
model name : Intel(R) Xeon(TM) CPU 2.40GHz
stepping : 7
cpu MHz : 2399.396
cache size : 512 KB
fdiv_bug : no
hlt_bug : no
f00f_bug : no
coma_bug : no
fpu : yes
fpu_exception : yes
cpuid level : 2
wp : yes
flags : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm
bogomips : 4784.12


p.s. thanks for the info - I will wait until the next WU arrives and see :)

[Ugly n Grey]

Sorry didn't know it was a Linux install :) I mean I knew the cluster was linux I just thought the workstation might be a dozy

You do need to increase verbosity (-verbosity 9) the final benchmarking number is missing from that log because of that. But Like I said, you should get a QMD packet next

[Ugly n Grey]

something interesting as a side note on Linux; some users claim to get a boost from running the Windows verrsion under Wine.... now why in the heck would that be? I only have one Linux box and it has no Wine install, I may try it.

[angra]

Sorry didn't know it was a Linux install :) I mean I knew the cluster was linux I just thought the workstation might be a dozy

You do need to increase verbosity (-verbosity 9) the final benchmarking number is missing from that log because of that. But Like I said, you should get a QMD packet next


2 more pieces of info. Reran one of the sessions with -verbosity 9, and here is the benchmarklines:


[13:56:37] + Benchmarking ...
[13:56:43] The benchmark result is 585
[13:56:43] - Preparing to get new work unit...


also, "free" reports enough memory free (>470MB at core selection time), but buffers are (as is normal) taking up as much as they can, so there actually only 200 or so MB reported as totally free. I don't know if the client is smart enough to add back buffers when determining free memory.

[Ugly n Grey]

Actually it looks for total memory, the client is not that smart. Even on a dual proc machine it never stops to consider that memory is shared. That benchmarking number is sufficient to get a QMD as far as my searching can tell. So lets see what happens :)

[Revv23]

Ok heres how i did it, get the 5.04 beta console, run the config, im not sure quite what the prompts were but it asks you something about scientific cores or large cores or something like that, basically thats the advanced methods command.

to do this youll need at least a gb of ram, as QMD's take up tons of memory (i havent noticed a performance hit in games, must page the info or something) the beta clint asks you how much ram you want to alocate to folding, make sure its set to at least 1000.

My first core was gromacs, i think it was more of a stability test/ benchmark of my rig, ive gotten two QMD's after my first gromacs core.

edit - make sure you are using -forceasm and -advmethods!!

[angra]

I will see how the next core is. It'll take a while to complete the first cores, will report back later.

can you guys tell it's a friday and I don't want to get back to the papers I need to finish today?

[dicecca112]

this a single instance of QMD? What are your system specs

[Ugly n Grey]

Read up, no QMD yet, waiting to get one :)

LOL Angra, you lazy bum. I have a client research document due today, I am feeling the same..

[dicecca112]

well in that case if you have the -advmethods flag and bigpackets=yes in the user.cfg file, will guarentee QMDS.

[Ackbar]

Hmm... I had no idea that F@H did semi-empirical or even quantum calculations. I thought it was only molecular dynamics. As far as the level of physics/science is concerned I feel better about putting my computers on F@H now rather than D2OL. I've always felt MD has been a little hokey for me to generally trust its results. I'm going to put two computers on F@H instead of just one now.

[Revv23]

Hmm... I had no idea that F@H did semi-empirical or even quantum calculations. I thought it was only molecular dynamics. As far as the level of physics/science is concerned I feel better about putting my computers on F@H now rather than D2OL. I've always felt MD has been a little hokey for me to generally trust its results. I'm going to put two computers on F@H instead of just one now.

Your going to need it good buddy, im about to pass you up :toast:

Seriously though, all in good fun, thanks for everthing you can spare :)


BTW angra my first gromacs core took an entire day to complete, now i put out a QMD in what looks like five hours less then that.

[angra]

BTW angra my first gromacs core took an entire day to complete, now i put out a QMD in what looks like five hours less then that.


excellent.

my dual-xeon box doing this is a bit outdated as you can see by the cpuinfo excerpt above. it might take more than a day. in fact, it's about 40% through after 8 hours so I guess it won't.

I will report back when it does anything interesting!

[Ugly n Grey]

I'm going to put two computers on F@H instead of just one now.

I feel that way as well :) Thanks for the boost Ackbar, we need it :) :toast:

[Ackbar]

Still no QMD here... personally, I'd rather only do QMDs if that were possible. On the FAQ they list that QMDs will be slower on AMD machines without using a hacked compiler... that sucks.

[dicecca112]

Oh Im sorry, I didn't realize Ackbar was on A64, with A64s you rarely get them because they are for SSE2 on Intels, you will get them occasionly, but Intel is guarenteed with advmethods and bigpackets=yes

[Ugly n Grey]

actually QMD requires dual precision FPU , which means SSE2, which means you won't get them ever on A64 unless it's a Boo Boo

[Ackbar]

Oh Im sorry, I didn't realize Ackbar was on A64, with A64s you rarely get them because they are for SSE2 on Intels, you will get them occasionly, but Intel is guarenteed with advmethods and bigpackets=yes

I'm also running on a Pentium-M that hasn't gotten a QMD yet either... only gromacs so far.

[Ugly n Grey]

Yeah, the Intel compiler used.... if it doesn't send back a correct CPU ID then no QMD's. Pentium M are also on the victims list it seems according to the reading. I have to wonder if the ID reported back by the compiler for Angra's Xeons on the workstation is also too far away from a standard P4... drag.

[angra]

both procs are on the second WU and both are doing another gromacs.

that is a good point to consider, UnG, about the Xeon vs. P4

[Ugly n Grey]

I'm kind of annoyed about the scoring actually. I'm not all that annoyed at Pande group, but I am really annoyed at Intel and their EULA for compilers.... I better rake someone at Intel over the coals before I lose my coporate phone number.

[angra]

so trying for after to restart until I got a QMD with no success, and after rereading this thread, doing some googling, and finding

http://forum.folding-community.org/viewtopic.php?t=13295

I decided to install wine on the machine, and run the windows console under wine.

got a QMD wu on my first try. go figure.

[Ugly n Grey]

It's funny you mentioned that, I had a big post ready to go on that topic (improved performance over Linux on Wine) but I hadn't posted it yet.

[angra]

It's funny you mentioned that, I had a big post ready to go on that topic (improved performance over Linux on Wine) but I hadn't posted it yet.

now I just need to figure out how to get wine working properly on the compute nodes. I fought with it a bit last night and made some progress at least getting the client running, but next I need to get the proxy working properly.

[Ugly n Grey]

that sounds like a cool experiment, let us know if we can assist with any specific problem.

[angra]

that sounds like a cool experiment, let us know if we can assist with any specific problem.


well the windows command line client does not seem to dump as much info in -verbosity 9 mode as the linux one, so it's a little trickier. by examining the proxy logs on the front node, the compute node client from within wine is not using the proxy configuration from the client.cfg. It is using the right user name and team number though. Anyone have a guess? Maybe the client.cfg file format is a little different under windows? Does anyone have a cfg file on a machine that must use a proxy to access the internet?

exposing the compute nodes directly to the internet is not an option.

[Ugly n Grey]

well the windows command line client does not seem to dump as much info in -verbosity 9 mode as the linux one, so it's a little trickier. by examining the proxy logs on the front node, the compute node client from within wine is not using the proxy configuration from the client.cfg. It is using the right user name and team number though. Anyone have a guess? Maybe the client.cfg file format is a little different under windows? Does anyone have a cfg file on a machine that must use a proxy to access the internet?

exposing the compute nodes directly to the internet is not an option.

Yes the windows version of the config file is formatted differently. Can you sent me you existing Linux file and I will convert it. Include your proxy information and I will add it to the client config. you can at least try it and once you see a sample will know immediately what has been done.

[angra]

Yes the windows version of the config file is formatted differently. Can you sent me you existing Linux file and I will convert it. Include your proxy information and I will add it to the client config. you can at least try it and once you see a sample will know immediately what has been done.

here it is. note the team is a temporary change for helping out Xciter.


[settings]
username=Angra
team=3074
asknet=no
machineid=2
local=43

[http]
active=yes
host=frontend-0
port=3148
usepasswd=no

[Ugly n Grey]

here it is. note the team is a temporary change for helping out Xciter.

Try this one :)

[angra]

Try this one :)

excellent, I think "usereg=no" was the magic button.

tweaking my setup/launch scripts now. I think that I have 24 CPU running QMD for team XS and 24 CPU running traditional for Xciter (for the moment).

gotta keep an eye on the temps for a little bit.

[Ugly n Grey]

I think that I have 24 CPU running QMD for team XS

:eek: :eek2:

Awesome. I have all these Opty's and they don't do squat despite the power because of the executable. I'm jealous. WTG :) :toast:

[Ackbar]

Just got my first QMD on a P4 laptop. This laptop is kinda old so it's going to take nearly 2 days to finish. Is there some way to trick my Pentium M (different computer) to recieving QMD jobs?

[angra]

Just got my first QMD on a P4 laptop. This laptop is kinda old so it's going to take nearly 2 days to finish. Is there some way to trick my Pentium M (different computer) to recieving QMD jobs?

It's been about 2 days and not a single one of my QMD units has completed yet!

I don't know if I am getting super-long WUs or if the wine performance is a hindrance or what.

we'll see if I end up getting 3k points for each unit or something :)

[Ackbar]

I'm curious, since F@H seems to be so strict about the processor ID and the jobs it gives it... will Pentium-D or Xeons get QMDs?

[Ugly n Grey]

well QMD's bring 450 points per unit. A 3.0 GHZ p4 with dual channel RAM can do a QMD in around 20 hours on Windows...longer on Linux... Wine is actually supposed to improve performance in some cases so go figure. I tried it on an Linux emulator on Solaris 64 bit and it sucked... two weeks to do a Gromac so i stopped ASAP.

Pentium D and Xeons bases on P4 type cores should get QMD's according what I read at the Stanford site, it just seems a lot of folks are using Winblows and not Linux ....

And Ackbar, I only wish there was a way to trick other procs into getting QMD's. But they need all the units worked on so it's not just about us scoring :) It's all good.

[Ackbar]

And Ackbar, I only wish there was a way to trick other procs into getting QMD's. But they need all the units worked on so it's not just about us scoring :) It's all good.

Hehe... that's true. The only reason I want QMDs over others is more due to my personal belief that QMDs being scientifically more meaningful.

I might try and pick up a dual Xeon setup, money permitting, to get more QMDs (and for some personal projects when I get there).

[Ugly n Grey]

Well what I did was resurrect a couple old P4's at 2.4 and 3.0 GHz respectively in order to play with QMD's/Linux combos. Cost me about 200 bucks in parts total. I already had HD's and cases/PSU's. The 3.0 GHz uses dual channel RAM... about 20 Hours per QMD, the 2.4 does not, about 2.5 days per QMD.... big difference

What I am finding is it's not the proc that is worked all that hard, it's the memory bus, if you have a faster memory bus these things crank FAST. Can you imagine with on one of my eight way rigs with a NUMA enabled OS and all the RAM kicking butt ? :slobber:

[Ackbar]

Well what I did was resurrect a couple old P4's at 2.4 and 3.0 GHz respectively in order to play with QMD's/Linux combos. Cost me about 200 bucks in parts total. I already had HD's and cases/PSU's. The 3.0 GHz uses dual channel RAM... about 20 Hours per QMD, the 2.4 does not, about 2.5 days per QMD.... big difference

What I am finding is it's not the proc that is worked all that hard, it's the memory bus, if you have a faster memory bus these things crank FAST. Can you imagine with on one of my eight way rigs with a NUMA enabled OS and all the RAM kicking butt ? :slobber:

Interesting find! Unfortunately, I have a bunch of older (when they first came out) P3's but no P4's sitting around (I didn't have money for a LONG time so I basically skipped a computer generation). I'm curious how a Xeon rig with some good ram will do then. I'm looking at some under $100 1.6LV Xeons right now... not sure if my budget permits if I have to buy a board, PSU...

[angra]

Well what I did was resurrect a couple old P4's at 2.4 and 3.0 GHz respectively in order to play with QMD's/Linux combos. Cost me about 200 bucks in parts total. I already had HD's and cases/PSU's. The 3.0 GHz uses dual channel RAM... about 20 Hours per QMD, the 2.4 does not, about 2.5 days per QMD.... big difference

What I am finding is it's not the proc that is worked all that hard, it's the memory bus, if you have a faster memory bus these things crank FAST. Can you imagine with on one of my eight way rigs with a NUMA enabled OS and all the RAM kicking butt ? :slobber:

are you doing it under wine or on the linux console?

[Ugly n Grey]

I'm playing around with it all Angra, about last week I identified a need for some solid research and answers before this team starts getting big (and it will, I promise). My problem is just googling it results in poor research, I'd like to have some certain answers.

[angra]

My problem is just googling it results in poor research, I'd like to have some certain answers.

The scientist in me is so proud :)

I am happy to help with the research front in any way I can. travelling a bit this week though.

[Revv23]

man i come home today and my pc went down at 4 in the morning on saturday night...

not sure if it was my roomate getting tired of my fans or she crashed... and damn i dont feel like going through the memory dumps...

[Revv23]

ok redid wiring and cut out the plastic holdng up the xp-120 on the retention bracket...

temps went from 55 stock loaded to 44 stock loaded...

hmm, feel like i still can get it lower, ive got screens of it maxed out at 1.725v 3.85 priming at 41c.... thats with fans on high, but with that kind of OC, i feel like i should be way below that at stock with fans on low.

[Ugly n Grey]

ON an Intel chip? Really? I don't think anyone else on air is doing any better really if they are in a case.

[Revv23]

well i was, so im going to keep working until i get back to it.

as i said i was low 40's loaded overclock intel on air.

[angra]

I think I was being hurt by disk performance - all the node clients were running out of NFS mounted directories.

I've switched my launch script to make everything run out of local scratch space. Will know in a couple of days if it helped. At least the load average on the NFS server is way down :).

[Ugly n Grey]

Interesting, I never would have even though of that. I'm so used to the SAN that I know I can't exceed it's capacity, but a straight NFS server or NAS system might get hurt pretty bad luanching all at once.

Still, once these things load, they are almost all in memory, writing to disk only when startup/shutdown or completion for send/receive occurs. So I'm not sure if you will see a performance increas. Please do let us know :)

[angra]

Still, once these things load, they are almost all in memory, writing to disk only when startup/shutdown or completion for send/receive occurs. So I'm not sure if you will see a performance increas. Please do let us know :)

I should say that the NFS load issues I was getting started when I converted to running within WINE.

I suspect there might be an issue specifically with WINE and disk access.

What I really want to know, though, is if this speeds up the work units. won't know that one for a while.

[mjgent]

These QMDs are great! I'm putting out about 1 per core in about 16hrs with an 820d o/c to 250fsb. That's about 1350 ppd on only 450pt QMDs. It's even better ppd if I get a few 600 pointers in there. These are coming in consistently for the past 3 days. :woot:

[Ugly n Grey]

very cool, what's your folding ID?

[mjgent]

very cool, what's your folding ID?

mjgent

[Ugly n Grey]

I can't see you in our stats?

[dicecca112]

that's because he doesn't fold your you guys http://folding.extremeoverclocking.com/user_summary.php?s=&u=105049

[Ugly n Grey]

LOL, no worries, as long as you participate, who cares :)

[mjgent]

LOL, no worries, as long as you participate, who cares :) When I started folding last year I hadn't found this site yet.

[bmac11]

How do I change bigpackets to yes? If I close the program and do it manually, I lose all the information in the cfg file.

[Ugly n Grey]

what version of the program are you running? and why would your client cfg file go away? tell me what you are doing to launch the application, maybe I can help. :)

[mjgent]

Yeah, you shouldn't loose you cfg file. There are a few ways to add bigpackets. You can add the flag to the target path of your folding shortcut or you can d/l the newest folding@home text only client which has this option during the setup.

[bmac11]

I'm running 5.03. Doesn't it update automatically?

If I go into the cfg and change the no to yes, I lose everything in the cfg file next time I start it up. It doesn't matter where I launch it from. It will do it if I launch it directly from the exe in the directory (gui or command mode).. I normally launch from a shortcut on the desktop and add the following to the target for both command and gui. "-local -advmethods -forceasm -verbosity 9." I run one of each bc of the P4's HT.

[Revv23]

check out ugly and grey thread, he has a good guide to running more then one instance, i had that same problem and it was due to setup issues, i also found that running just one instance on ht makes your computer feel faster and it doesnt fold too much less.

[mjgent]

It will update the next protein to fold but not the program. You'll need to save over 5.03 with 5.04 that just came out. 5.04 has the ability to config big packetts and advmethods in the config process. You might want to start with a fresh install.

I don't want to assume but are you running 2 clients from different folders (directory) and each one being configed with differnt machine IDs?

[bmac11]

They are both set to machineid=1. Even if I try to change the machineid in the file, I will lose all the information in it next time I start it up.

Yes, they were installed in different directories.

Could I just reinstall and then copy over "work" directory?

[mjgent]

You will need to set the Machine IDs as 1 & 2 not both 1. I'd start over with clean folders and have the client redownload everything again after you set the Machine IDs correctly.

[Ugly n Grey]

I agree, I would start fresh and edit out any registry tags if you are running it as a service as well.

[bmac11]

Do I have to edit the cfg file to change the maching id? Is that the only way? I just think that if I reinstall everything, it will act the same. I guess I'll try tomorrow when the WU finishes.

[Ugly n Grey]

you can run the application with the -configonly flag and just configure it again for a different machine ID

[mjgent]

I agree. It's too easy to mess up the cfg file directly. Use the -config only flag and set everything from there.

[bmac11]

word... I never knew about that flag.. problem solved!

back to folding!
thanks guys

...hopefully I will have vapo lightspeed this week and I will build an opteron 939 system with it within the month.. maybe dual core. That should produce some numbers.. every box counts.

[mjgent]

Nice going. That should be a nice rig! :slobber:

[bmac11]

One more thing.. what are these in the command console version.. "Finished a frame (12)".. and they repeat.. I've never seen that before. Different WU bc of the changes I made?

[mjgent]

One more thing.. what are these in the command console version.. "Finished a frame (12)".. and they repeat.. I've never seen that before. Different WU bc of the changes I made?

A frame is just a section or step of the WU (work unit). One WU is a protein that is downloaded and "worked on" or folded and then the results are uploaded to the folding@home servers when complete.

[bmac11]

I know.. I was just wondering bc its counting them differently.. one by one instead of in a large group.. oh well.. as long as I get some pts.

[mjgent]

That's normal for the text client.

[GoodOmens]

actually QMD requires dual precision FPU , which means SSE2, which means you won't get them ever on A64 unless it's a Boo Boo

Dont the new AMD's support SSE2? Mine says it does in cpu-id

[Ugly n Grey]

Yes the new AMD CPU's do support it. The problem is the compilers they used from Intel connaot identify the capabilities of non AMD CPUs...usual Intel anti competitive stuff....

[GoodOmens]

Yes the new AMD CPU's do support it. The problem is the compilers they used from Intel connaot identify the capabilities of non AMD CPUs...usual Intel anti competitive stuff....


Ok that makes sense ... stupid intel :slapass:

Guess I'll just use my machine at work for folding ... just set it up and got a qmd on first try.

[GoodOmens]

Yea my 3 ghz p4 at work really cranks them out :-)

[angra]

I really wish they'd fix the linux client so that it gets QMDs on intels. I am getting at most ~150PPD on my xeon CPUs under WINE (3 GHz), and I want the performance you guys are getting :)

[da-key]

I am getting great points per unit. Think its because of these QMD's and the Intel.

[Revv23]

I really wish they'd fix the linux client so that it gets QMDs on intels. I am getting at most ~150PPD on my xeon CPUs under WINE (3 GHz), and I want the performance you guys are getting :)


yeah that sucks, my moms 2.0ghz northwood does better then that... it pulls about 450 in a day and a half.

[Ugly n Grey]

I'd have to agree da-key, you are racking up the points. You've made it up to 20th place already in a short time.

Angra. I have the following definite information.
Linux and Windows operate at roughly the same speed without WINE. (at least most Debian distros)
The Linux assignment server however has less QMD cores to hand out and you are less likely to get one.
Running the client on WINE turns it to the Windows assignment server where you are more likely to get a QMD packet.
Running WINE can slow your performance up to 2 or 3 times over native Windows or Native Linux....

I have tested all of this... WINE gets you QMDs all the time but slows performance...In native mode the machine can already do 150-200ppd or thereabouts without WINE and roughly the same with WINE and QMD's. Therefore, WINE is a hassle that gains you nothing really in terms of RAW production.

AMD cores can get an occasional QMD packet on Linux from some reports. I did not get one in my limited testing.

[angra]

I have tested all of this... WINE gets you QMDs all the time but slows performance...In native mode the machine can already do 150-200ppd or thereabouts without WINE and roughly the same with WINE and QMD's. Therefore, WINE is a hassle that gains you nothing really in terms of RAW production.


Interesting - I dont think I was getting that rate of production in native linux - more like ~100-120. I was averaging around 10k per day with 88 CPU. Though, it is possible that my results were skewed somewhat by having differing types of CPUs online. Maybe the Opterons were lagging and the Xeons really were getting 180-200.

Anyway - thanks for the digging, this is interesting stuff.

I really want to build a cluster of GPU enabled workstations for testing out distributed GPGPU type work - once the Stanford guys release their GPU enabled cores, that could really fly :).

But...you know...clusters aren't free :)

[mjgent]

I am getting great points per unit. Think its because of these QMD's and the Intel.I am really starting to like these Intel dual cores with QMDs. These things are pumping out the points. If you jack up the fsb on these you can run 2 QMDs at once (each core) and they seem to do well. I can't wait for the new dual cores coming out. :D

[nhenhophach]

I am really starting to like these Intel dual cores with QMDs. These things are pumping out the points. If you jack up the fsb on these you can run 2 QMDs at once (each core) and they seem to do well. I can't wait for the new dual cores coming out. :D
I think you can get ~ 1000ppd.

[mjgent]

I think you can get ~ 1000ppd.Yep. You're just about right. :D

[nhenhophach]

He he i have use it.

[Ugly n Grey]

But...you know...clusters aren't free :)

LOL, no they are not free and I have built a few of them. Though at times I got too lazy and just ordered a superdome. It easier to configure :) and pretty much guaranteed to work. The HP Superdome is also the only truly sucessful implementation of the Itanic (Itanium) processor family in my opinion.

Angra, do you have anyway of testing memory bandwith on those Xeons? I mean that is the number one factor defining completion rate of these QMD's. For example, a single core P4 with a gig of single channel RAM can take three days to do a QMD at 2.4GHz. That same processor with dual channel can do that QMD in about 32 hrs or less depending.

[angra]

Angra, do you have anyway of testing memory bandwith on those Xeons? I mean that is the number one factor defining completion rate of these QMD's.


I'm sure I do. Any particular benchmark in the linux space that is considered authoritative? If nothing else I can code up a test.

[angra]

Attached is a zipped output from the program "cachebench" which theoretically tests cache and memory bandwidth.

[Ugly n Grey]

Sorry Angra, that's a CPU cache intensive testing program we are sort of looking for sustained throughput. You can use RAMSPEED http://duplex.hypermart.net/ramspeed/ It's quick and easy.

If you feel like you want a little more insight into your systems power also check out this nifty little tool :D . I have played around with this, but claim no great knowledge.
http://www.cs.virginia.edu/stream/

Cheers
UnG

[angra]

OK - here's a set of ramspeed results...

[Ugly n Grey]

OK - here's a set of ramspeed results...

Writing on the larger blocks looks slow....but over all not bad, the larger L2 Cache on the Xeons really shines. Here is an output from a basic P4 at 3GHz with DDR 200 RAM for comparison, below is the AMD results at DDR 260...

Basically I am saying nothing really seems amiss... I just don't understand it. Pythagoras may have some contacts in the same boat as you are in to help squeeze the most out of your boxes...

Intel P4, 3.0 GHz, DDR at 200 MHz
A:\ramspeed>ramspeed -b 1
RAMspeed (DOS) v2.3.2 by Rhett M. Hollander and others, 2002-05

4Gb per pass mode

INTEGER & WRITING 1 Kb block: 5572.79 Mb/s
INTEGER & WRITING 2 Kb block: 5425.17 Mb/s
INTEGER & WRITING 4 Kb block: 5527.67 Mb/s
INTEGER & WRITING 8 Kb block: 5439.58 Mb/s
INTEGER & WRITING 16 Kb block: 5572.79 Mb/s
INTEGER & WRITING 32 Kb block: 5454.06 Mb/s
INTEGER & WRITING 64 Kb block: 5550.14 Mb/s
INTEGER & WRITING 128 Kb block: 5417.99 Mb/s
INTEGER & WRITING 256 Kb block: 5454.06 Mb/s
INTEGER & WRITING 512 Kb block: 4970.87 Mb/s
INTEGER & WRITING 1024 Kb block: 1667.75 Mb/s
INTEGER & WRITING 2048 Kb block: 1668.43 Mb/s
INTEGER & WRITING 4096 Kb block: 1652.28 Mb/s

A:\ramspeed>ramspeed -b 2
RAMspeed (DOS) v2.3.2 by Rhett M. Hollander and others, 2002-05

4Gb per pass mode

INTEGER & READING 1 Kb block: 10981.23 Mb/s
INTEGER & READING 2 Kb block: 11473.39 Mb/s
INTEGER & READING 4 Kb block: 10778.95 Mb/s
INTEGER & READING 8 Kb block: 10694.52 Mb/s
INTEGER & READING 16 Kb block: 7787.07 Mb/s
INTEGER & READING 32 Kb block: 7801.90 Mb/s
INTEGER & READING 64 Kb block: 7787.07 Mb/s
INTEGER & READING 128 Kb block: 7801.90 Mb/s
INTEGER & READING 256 Kb block: 7641.79 Mb/s
INTEGER & READING 512 Kb block: 7098.79 Mb/s
INTEGER & READING 1024 Kb block: 3268.95 Mb/s
INTEGER & READING 2048 Kb block: 3300.56 Mb/s
INTEGER & READING 4096 Kb block: 3271.57 Mb/s

AMD a64 at 2.6GHz DDR at 260MHz

A:\ramspeed>ramspeed -b 1
RAMspeed (DOS) v2.3.2 by Rhett M. Hollander and others, 2002-05

4Gb per pass mode

INTEGER & WRITING 1 Kb block: 16449.80 Mb/s
INTEGER & WRITING 2 Kb block: 16583.00 Mb/s
INTEGER & WRITING 4 Kb block: 16253.97 Mb/s
INTEGER & WRITING 8 Kb block: 16516.13 Mb/s
INTEGER & WRITING 16 Kb block: 16583.00 Mb/s
INTEGER & WRITING 32 Kb block: 6005.87 Mb/s
INTEGER & WRITING 64 Kb block: 16516.13 Mb/s
INTEGER & WRITING 128 Kb block: 23141.24 Mb/s
INTEGER & WRITING 256 Kb block: 6714.75 Mb/s
INTEGER & WRITING 512 Kb block: 6714.75 Mb/s
INTEGER & WRITING 1024 Kb block: 2244.38 Mb/s
INTEGER & WRITING 2048 Kb block: 2216.45 Mb/s
INTEGER & WRITING 4096 Kb block: 2210.47 Mb/s

A:\ramspeed>ramspeed -b 2
RAMspeed (DOS) v2.3.2 by Rhett M. Hollander and others, 2002-05

4Gb per pass mode

INTEGER & READING 1 Kb block: 16516.13 Mb/s
INTEGER & READING 2 Kb block: 16583.00 Mb/s
INTEGER & READING 4 Kb block: 16062.75 Mb/s
INTEGER & READING 8 Kb block: 16384.00 Mb/s
INTEGER & READING 16 Kb block: 16583.00 Mb/s
INTEGER & READING 32 Kb block: 16650.41 Mb/s
INTEGER & READING 64 Kb block: 16583.00 Mb/s
INTEGER & READING 128 Kb block: 8770.88 Mb/s
INTEGER & READING 256 Kb block: 8497.93 Mb/s
INTEGER & READING 512 Kb block: 8677.97 Mb/s
INTEGER & READING 1024 Kb block: 3567.94 Mb/s
INTEGER & READING 2048 Kb block: 3567.94 Mb/s
INTEGER & READING 4096 Kb block: 3465.31 Mb/s

[mjgent]

Writing on the larger blocks looks slow....but over all not bad, the larger L2 Cache on the Xeons really shines. This is good news for the Presslers too.?

[Ugly n Grey]

The presslers do fine with QMD's especially with the DDR 2 RAM.. memory bandwidth makes a difference, that's for dure....

[Revv23]

just for reference, according to FAH mon my mom's 2.0ghz P4a 400mhz FSB with 512mb single channel DDR400 takes 34.5 minutes per frame or 2.36days per 450 point QMD.