[Shaktai]

F@H -Gromacs Core for Mac, Supposed to be available today. Here is the thread in the Folding@Home forum

http://forum.folding-community.org/v...37&start=0

I suspect the CLI version might be required. Gromacs takes advantage of the G4's Altivec. Macs are about to get "competitive on this project. Also, the time stamping issue is supposed to be fixed.

Oh, be sure to look for version 3.25, NOT 3.24.

[reader50]

The client downloads PPC cores from here:

www.stanford.edu/~pande/OSX/PowerPC

You can already download the two core78 Gromacs files. In theory, all we need is for the Assignment Servers to start assigning Gromacs units to OSX. Just in case, I have had the "-advmethods" tag set on my OSX client for some weeks now. The tag is accepted but so far has done nothing.

[mikkyo]

And there are 2 files there from yesterday.
Core_78.exe.bz
Core_78.fah
Now what?

[reader50]

I suspect that you need both for gromacs. In the meantime, from the linked thread, you can download the 3.25 OSX clients from this page:

folding.stanford.edu/OSX/325.html

According to the Pande Group posts, use the "-advmethods" to try and get gromacs units. You may not get gromacs units right away, but the new clients will log much more activity.

The CLI client download is the client itself only. Unstuff, chmod +x, stop existing Fold CLI client. Replace the client file (~262 KB filesize) and restart.

[enola]

So just replace the new client with old right? All the existing support files and work directory can remain untouched? Very excited to see how gromacs does in a G4. I think this Friday my Mac will turn in it's current work unit so I'm crossing fingers.

Edit: Oops, reader50, just saw your post. Got it. Just replace the client.

[Shaktai]

And wouldn't you know that my PowerMac just started a 40 pointer. Going to be 5-6 days before I can even get a chance to play with
Gromacs.

On the upside, upgraded the CLI on both my macs and restarted with the -advmethods flag. Really like the time stamping on it now. Changed over when I came home for lunch, and now that I am off work, I can see right in the terminal how many frames are completed and how long it has taken per frame.

iBook 600, just under 17 minutes per frame on a P633. Now if I can just catch some Gromacs one of these days. Looks like server .114 is dishing out OS-X gromacs now.

[DAlex]

Originally posted by reader50:
Just in case, I have had the "-advmethods" tag set on my OSX client for some weeks now. The tag is accepted but so far has done nothing.

Where do I need to add the -advmethods text? Is it in the client config? Sorry if this is a daft Q

DAlex

[Shaktai]

Originally posted by DAlex:
Where do I need to add the -advmethods text? Is it in the client config? Sorry if this is a daft Q

DAlex

For now, you need to use the CLI client and start it up adding -advmethods after the console name in the terminal like so....

./OSX-3.25 -advmethods

[DAlex]

Originally posted by Shaktai:
For now, you need to use the CLI client and start it up adding -advmethods after the console name in the terminal like so....

./OSX-3.25 -advmethods

Cheers, I added it to my startup script,
#!/bin/sh
./folding -local -advmethods

Appeared to load OK, just got to wait for it to finish the current unit to see if it gets a gromacs.....

DAlex

[Shaktai]

Originally posted by DAlex:
Cheers, I added it to my startup script,
#!/bin/sh
./folding -local -advmethods

Appeared to load OK, just got to wait for it to finish the current unit to see if it gets a gromacs.....

DAlex

UPDATE: From the F@H forum.

It's likely that the new core will be beta tested for "a day or three" before it is released for those that specify advmethods. If all goes well in early beta testing, the next level of testing will be advmethods. If all goes well in advmethods, it will then move to into general distribution.

If you aren't on the "beta" team, they you won't get them yet. The major bugs were caught during Windows and Linux testing, so this phase is expected to go quickly. -- I am anxious to hear from anyone on Beta Team how they do.

[bousozoku]

I downloaded the new client this morning. Of course, this was right after it had uploaded a work unit.

Everything seems to be running nicely and there's more information about the current progress in each foreground process.

Have to wait until later to see if it downloads a new core to go with the a Gromacs WU. As well, as much as my Gromacs processing code worked for my progress monitor. I wasn't able to live with it day-to-day, so I'll be happy to see a further validation of that work.

[reader50]

I did some processing of the Gromacs benchmarks chart. Converted it to Score per MHz, then normalized it to the single-CPU G4, and sorted by the avg Score/MHz for each test system:

Code:
.....CPU ....... MHz	Vill	L/C.	L/P.	DPPC	P-CH	Avg.
.1 IBM Power 3... 395	1.06	1.12	1.45	1.68	1.95	1.45
.2 SGI R12000.... 270	0.83	0.88	1.68	1.35	2.23	1.39
.3 Alpha 21264.. 1000	0.87	0.90	1.55	1.25	1.80	1.27
.4 Itanium....... 800	0.63	1.22	1.30	1.00	0.91	1.01
.5 PPC G4....... 1250	1.00	1.00	1.00	1.00	1.00	1.00
.6 Alpha 21264... 667	0.62	0.71	1.25	0.97	1.35	0.98
.7 Pentium 3.... 1133	0.71	0.76	1.03	0.88	1.09	0.89
.8 Athlon....... 1533	0.71	0.68	0.88	0.72	1.11	0.82
.9 Athlon.... DP_1533	0.64	0.69	0.82	0.98	0.83	0.79
10 PPC G4.... DP_1250	0.73	0.81	0.79	0.78	0.71	0.76
11 Ultrasparc IIi 440	0.50	0.51	0.83	0.77	1.14	0.75
12 Xeon P4...... 2200	0.63	0.67	0.78	0.67	0.83	0.72
13 Pentium 3..... 800	0.57	0.60	0.83	0.76	0.79	0.71
14 Athlon........ 800	0.60	0.56	0.83	0.69	0.83	0.70
15 Xeon P4... DP_2200	0.55	0.62	0.68	0.88	0.68	0.68
16 Pentium 4.....1700	0.61	0.59	0.73	0.67	0.83	0.68
17 Pentium 4.....2000	0.61	0.54	0.68	0.62	0.83	0.66
18 Pentium 3.. DP_800	0.50	0.50	0.55	0.87	0.57	0.60

[Shaktai]

Originally posted by reader50:
I did some processing of the Gromacs benchmarks chart. Converted it to Score per MHz, then normalized it to the single-CPU G4, and sorted by the avg Score/MHz for each test system:

Hmm! Something looks strange about those scores, but I am too tired to figure it out. Something just doesn't seem to mesh though. Have to look at it again tomorrow I guess, when I am rested.

[mikkyo]

Originally posted by Shaktai:
If you aren't on the "beta" team, they you won't get them yet.

Well like they should at least turn them on for advmethods, I mean you have to specify that.
Beta is over! Gives us the goods!

[Shaktai]

Originally posted by mikkyo:

Well like they should at least turn them on for advmethods, I mean you have to specify that.
Beta is over! Gives us the goods!

Ask and you shall receive. Just posted today.

Hi,
Gro units are trickling out on advanced methods -- if you have the -advmethods flag there's a chance you may get one on an assignment today. Next week, we'll set it up so that all assignments to people using -advmethods flag will be Gro.
Thanks,
Guha

[enola]

Originally posted by Shaktai:
Ask and you shall receive. Just posted today.

Damn, I didn't get one when cpu#2 returned a work unit this morning. #1 will return one tonight so I still have another chance today. What's your project number? I just looked at the summary page and 656 and 657 are worth a whopping 146 points!

Edit: Duh! I just re-read your post. I interpreted your reply to mean you'd actually gotten one.

[jarling]

Is this -advmethods thing available only to those running the command line version? How would I do this if I'm running the GUI version?

[Shaktai]

Originally posted by jarling:
Is this -advmethods thing available only to those running the command line version? How would I do this if I'm running the GUI version?

Yes, right now it requires the CLI version. Soon, all clients should have access though is the last word I saw.

by Enola
Damn, I didn't get one when cpu#2 returned a work unit this morning. #1 will return one tonight so I still have another chance today. <edit> I just looked at the summary page and 656 and 657 are worth a whopping 146 points!

I haven't gotten one yet either, but hopefully will when the current "long" tinker cores are finished. As "G" said, they are trickling out. However once the servers are configured next week, then folks running the -advmethods flag, should get Gromacs as long as any are availalbe.

Wow, those new proteins are something else, and they ARE Gromacs. It looks like this project is now going after some serious heavy duty proteins. From the Science standpoint, that is great, but from the participants standpoint we will have to continue to "be patient", comfortable with the idea that we are now doing hundreds of times the work we were a few months ago.

[jarling]

Has there been any talk in their forums about reducing the number of timesteps for the larger proteins. For instance, I'm running one right now that will take me a week that has 400 timesteps. If it has 100 I could be done in a few days. After all, the whole point of distributed computing is taking a large problem and chopping it up into smaller easier to manage pieces.

[reader50]

Originally posted by jarling:
Has there been any talk in their forums about reducing the number of timesteps for the larger proteins. For instance, I'm running one right now that will take me a week that has 400 timesteps. If it has 100 I could be done in a few days. After all, the whole point of distributed computing is taking a large problem and chopping it up into smaller easier to manage pieces.

That would be a smaller protein. You are assigned smaller proteins based on the benchmark value when the program starts up. If you want smaller proteins, always start Folding while launching the Classic environment.

In general, more steps are preferable because less work is lost when you stop and restart the client.

[bousozoku]

There were plenty of those 100 frame work units when I started with folding. They were almost all worth 0.6 points. My G3/400 was getting them done is several hours.

How long would it take to make a difference if you were only folding 0.6 point WUs?

More WUs, fewer points, but faster. It sounds like you'd be quickly going nowhere.

[Shaktai]

The protien sizes has always been a concern, even among project operators, as well as the performance disparity between platforms with Tinker.

That was one of the reasons for the Gromacs core, and all the work that has gone into it. It handles the larger proteins better with greater speed allowing them to tackle much more difficult possibilities. It will hopefully also restore parity among the various platfoms and I think is open source, which may help to speed up its advancement.

There will still be a tinker core for sometime, but the more complex proteins will be Gromacs. I suspect, that the reluctance to spend time, money and resources on improving the tinker core for platforms other then Windows, was because Gromacs was "just around the corner".

We will know in a few weeks time if gromacs lives up to its promise. Remember that first 70 or 73 point protein with the Tinker core? It had a 50 day deadline. The current 73 point gromacs core protein has a 20 day deadline. I have configured not only my G4 for advance methods, but also my G3 iBook, as I am anxious to examine the differences. Soon...very soon...!

In the early days of this project, we were really just working to "confirm and refine" the science and process to prove its validity against known targets. Now we are starting to move into the realm where "gradually our work will be one of "discovery", advancing the science and truly helping to find cures or even better, preventive treatments.

Technology moves quickly, but science requires a constant and steady pace. Much of scientific study today is focused not just on the immediate present, but also on laying the groundwork for future generations to build on.

[mikkyo]

Well all I can say is if the gromacs core lives up to the claims, TeamMacNN will reap the rewards, and you all are in serious trouble.
I don't have all my machines running, nor do the other big hitters. We all have been waiting on some project that will actually use the cpu power(and altivec) instead of just wasting it on non-optimized (non-fpu even!) functions.
So if gromacs actually puts us on par with the athlons, we will probably take the time to set everything up on F@H.
Then the battle for 1st will be between cobramac, macgeek4u, and myself.

So overall it looks like the team should see 30% gain with the new core, plus another oh 80GHz coming online.

[Shaktai]

Originally posted by mikkyo:
So overall it looks like the team should see 30% gain with the new core, plus another oh 80GHz coming online.

Well let's all hope that Gromacs lives up to its promise. Another 80 ghz huh? Wow!

[OneMacGuy]

Originally posted by mikkyo:
Well all I can say is if the gromacs core lives up to the claims, TeamMacNN will reap the rewards, and you all are in serious trouble.
I don't have all my machines running, nor do the other big hitters. We all have been waiting on some project that will actually use the cpu power(and altivec) instead of just wasting it on non-optimized (non-fpu even!) functions.
So if gromacs actually puts us on par with the athlons, we will probably take the time to set everything up on F@H.
Then the battle for 1st will be between cobramac, macgeek4u, and myself.

So overall it looks like the team should see 30% gain with the new core, plus another oh 80GHz coming online.

Not to mention little old me! I am just waiting for things to settle down so I can move all of my CPU's over to a worthwhile project. I am tired of looking for aliens that I don't believe exist anyways!!

[Shaktai]

Originally posted by OneMacGuy:
Not to mention little old me! I am just waiting for things to settle down so I can move all of my CPU's over to a worthwhile project. I am tired of looking for aliens that I don't believe exist anyways!!

Looks Like I might be hard pressed to stay in the top 5 before long.

[mikkyo]

Now, now, remember, you get an increase due to gromacs too.
I imagine you'll be up there for a while.
I'm still not convinced yet.
I haven't seen any of my macs processing with a gromacs core.

[Shaktai]

Originally posted by mikkyo:
Now, now, remember, you get an increase due to gromacs too.
I imagine you'll be up there for a while.
I'm still not convinced yet.
I haven't seen any of my macs processing with a gromacs core.

Heh heh! Yeah, but I got no reserves left for now, and my daily production is already #5.

Of course I am saving up for late summer, just in case Apple delivers something real nice, new and fast.

[reader50]

Stanford has adjusted the server. If you specify -advmethods now, you will get a Gromacs WU.

I just got one. Core_78 required. It's Project 908, 250000 steps. That project is not listed, but due date is 15 days. For Gromacs units, that looks like a 33 pointer.

[bousozoku]

Originally posted by reader50:
Stanford has adjusted the server. If you specify -advmethods now, you will get a Gromacs WU.

I just got one. Core_78 required. It's Project 908, 250000 steps. That project is not listed, but due date is 15 days. For Gromacs units, that looks like a 33 pointer.

I also just got one of those.

[reader50]

Update: p908 is a 100-frame, and I've completed one frame. About an hour, which translates to a 4-day run time for 33 points, or about 8 points per day. This compares to my 2 point per day avg recently on Tinker.

Preliminary conclusion based on 1 completed frame of p908:
PPC Gromacs is 4x faster

Edit: our team output hasn't spiked yet. The fastest PowerMacs should start finishing gromacs units late tonight or tomorrow, at which time we will see just how high we can go.

[mikkyo]

None of my machines have received any.
la la la la
Once you get them you are looking at what, 4 days, before you will notice the gain.
Hmm. End of the week should be quite exciting.

[reader50]

They have added p908 to the projects page - it wasn't there when I first got the unit. It is only a 13-pointer, with a 12-day completion (my client indicated 15 days allowed).

3 frames completed. Run time is 66 minutes if I use the computer for other things, 57.5 minutes if the computer does nothing else. At current rates it will complete in 4 days giving me a 3.25 per day avg, or just 62% faster than Tinker. Perhaps it will speed up in later frames.

My system is a G4 7400 at 350MHz (100MHz bus): 9.28 Score/Day/GHz.
A G4 7410 should score identically if the bus speed is the same.
G4 744x and 745x have additional Altivec power + faster bus speeds. Rate should be higher.
G3 rates -- no idea.

[bousozoku]

I'm seeing about 33 minutes per percentage point or 3300 minutes for the same 908 work unit, so a little over two days total with each G4/800.

[reader50]

I just tried running two clients on my single-CPU system, to see if I could get around whatever the PPC gromacs bottleneck is. Duplicated the client folder, changed the machineid to "2" on the copied folder, and started it.

Frame time doubled to ~2 hours per frame, up from one hour. That didn't work, but did not seem to hurt either, so it remains a useful trick for buffering Fold units.

Tonight I'm going to either log into >console or boot into single-user mode, and leave Folding running to see how much the frame time can be pushed down.

[enola]

I've got p907_villin_fo running about 32.5 minutes per frame. I've got a dual 1 gig. Don't know the exact Motorola model number for the processor but it's got 256K L2 cache and 2MB L3. Bus is 133MHz.

[kman42]

I just finished a WU and the new one didn't come in as gromacs. I have -advmethods flagged and the new client.

kman

[mikkyo]

Well I finally got one, a 908.
Looks like about 27 Hours to complete on one processor of a Dual 1.25Ghz G4.
The other processor has a Tinker.
Eh. 13 points huh?
Whee, 13 points a day.
I think my Althons are banging out a lot more than that.

[Shaktai]

I really think that something is not right about the 906-908 proteins or their scoring. Still trying to get some better answers over in the FAH forum. Even Guha states there should be about a 40% increase with altivec.

[bousozoku]

Originally posted by mikkyo:
Well I finally got one, a 908.
Looks like about 27 Hours to complete on one processor of a Dual 1.25Ghz G4.
The other processor has a Tinker.
Eh. 13 points huh?
Whee, 13 points a day.
I think my Althons are banging out a lot more than that.

If that 27 hours is right, your 1.25 GHz is about twice as fast as my 800 MHz which is running to 55 hours.

[Shaktai]

Originally posted by bousozoku:
If that 27 hours is right, your 1.25 GHz is about twice as fast as my 800 MHz which is running to 55 hours.

hmm! my 800 (single) with 256 L2 cache only, (no L3) is about 44 hours, running without interruption. You sure you don't have something else running in the background?

[mikkyo]

Different processor cores could explain the differences.
Which machines are you guys comparing?

As to the 40% increase, well we never ran a tinker project with gromacs, so maybe the 40% is just a tinker time/frame/gromacs time/frame.
Aren't the gromacs proteins more steps?

My Athlons(1900) seem to do about 5000 steps in 7 minutes running the gromacs core.

I bet a 3000 XP flies.

[bousozoku]

Well, there are no frames used in the Gromacs logging. There were steps in Tinker logging but we weren't really looking at it that way. From what I remember of all my machines had done, there were anywhere from 200,000 to 5,000,000 steps. So, I'd say there are fewer steps involved but then these Gromacs WUs have small point values.

As far as what I've got running in the background, there's not much other than the Intego NetBarrier firewall and the Palm desktop, but I do use the machine for everything.

The 908 I had concluded just a few minutes ago, so it took approximately 47.5 hours instead of 55.

To the point about x86 vs. PPC: it could be a couple of things, but it may come down to the data being structured in x86 format and our PPC machines, as usual, are being made to convert it. There is a significant I/O component in all of this as I've noticed that faster hard drive throughput helps. e.g., on Tinker, my dual G4/800 per processor is not twice as fast as my G3/400.

[mikkyo]

Frames, steps whatever they want to call it.
They probably did, time to complete/time to complete(gromacs) = big gain.
Woo woo.
What time do the stat servers update?
I think I will dump a couple of completed gromacs later today and want to catch the change.

You should run top -ud in Terminal to see who is eating your CPU time. I think you will find the Palm Hotsync process and Palm Desktop are little hoggish.
Are you running FAH at low priority?
If it were just disk access, then a RAM Disk would help.
If your hard drive is set to sleep, then I could see you losing some time to it spinning up ever time FAH goes to update its stuff, because it probably just went to sleep right before then.
Otherwise it really shouldn't be that disk intensive. I would bet Cache, and RAM speed have more to do with it. But if you two are running the same 800MHz machine, then it must be something else.

I haven't taken the time yet to run Shikari on the new FAH core, but it will be interesting to see what it spends most of its time doing.

Sometime in the next day our team should see the gains show up. We shall see how we stand then. Remember the other mac teams will have a gain too. It really is too bad we can't all be on the same team.

It would also be nice to have an all-machine monitoring app that didn't suck up CPU getting updates from all of ones machines. Something like the hardware monitor for Xserves that could report back the current progress of everything running.
Maybe something that servers up a simple version of the progress over HTTP and the monitoring client could have a list of all the IPs to hit and grab the progress every 20 minutes or something.
If you symlink the unitinfo.txt file to your Sites directory you can kinda have something like it now. Hmmm. Maybe it wouldn't be to hard to whip something up.

It would be much better to have it integrated into the FAH client though.
An option to turn on or off http output on some uncommon port. The monitor could just hit each ip at that port and get back a protein project number and a progress packet.

[bousozoku]

I'm already working on extending my monitor to work cooperatively with itself across a network.

A person will be able to specify a description, port, and master client IP address on each client. Each client would then report availability on information to the master client, which could then talk to the others and receive the data for display.

As far as the current speed of Gromacs for me, I'm not dissatisfied, just surprised. If it were 100 percent dedicated to folding, it would definitely be upsetting, but as I use it from time to time for internet business, programming, and general stuff, it is acceptable. I will try to disable the Palm pieces. I've left the folding priority at the default setting.

I received my first 908 on my G3 today, so we'll see how quickly that goes. It would be nice to say that each of my G4s are going twice as fast as my G3 is. [EDIT] It took 94 minutes for the first percent, which makes me feel much better about the G4. [/EDIT]

[mikkyo]

The key with getting the data off all your units is to use the lowest amount of processor power possible, yet get the needed data.
I started to play with a shell script that would pack the important stuff into an html file every 20 minutes or so and then a client could request that file from each machine.
The CPU usage spike is about half a second, so it shouldn't be too bad.

If you are going to have an app running on each machine, then it had better sit real idle until a request packet comes in, oh you are planning on pushing the data.

I wonder if there would be much problem if they all pushed at once.

[bousozoku]

It will use a queue to hold the requests. When each queue entry is serviced, the master client will request the data and it will be pushed toward the master client so it can update its information for the slave client.

[mikkyo]

How is that new version of the monitor coming along?

On another note, yesterday I had several machines with gromacs cores complete, yet fail to return and resume thier work.
They were sitting idle with no FAHCore running.
I looked at the logs and it would connect and request and receive a new work queue and then nothing.
I wondered if one of the servers was down over at F@H or something.
I sure hope the clients are well designed enough to recover.
I tried forcing a send all and it worked but when I relaunched the clients they still were hung at the same place.
So there I left them and went to bed.
It looks like they did recover.
I hope that this sort of thing doesn't happen too often.

[reader50]

All the PPC Gromacs units are coming from .114, and the time is about right for the first Mac gromacs adopters to be dumping. I'm at 99%.

Looking at 114 on the status page, it looks like it's under heavy load. There was a complaint in the project forum about 114 earlier today.

[bousozoku]

Yeah, .114 was under a big load last night that kept getting worse. It took 20 minutes to upload my 907 and I gave up after a total of two hours uploading and downloading without almost anything to show for it. I removed the -advmethods flag and got a Tinker WU within seconds of the change.

They probably saw the stats which show that there are fewer than 300 Macintosh folders and decided to give us a very small server.

As far as the monitor goes, I'm not that far yet. I just cleaned up the Gromacs estimates and uploaded the new version yesterday.