Reliable folding and overclocking article

[Xilikon]

After experiencing some hair-pulling issues with my main PC, which is in the end due to some overclocking settings slighty too much on the borderline of stability, I found out there are no reliable informations on how to tweak the overclocking enough to help make folding faster while eliminating all the issues which cause folding to fail to complete a WU or other issues.

We all know that when we overclock a computer, we are raising compontents in a computer over the design specifications to make it run faster. The usual components we overclock is :

-CPU : This will make calculations much faster.
-Memory : This will help improve the memory reads/writes so we can feed the CPU with data fast enough to not cause slowdowns while waiting for infos to be retrieved or written.
-Graphics card : this will help render the images (both 2d and 3d images) faster and reduce chopiness and make it able to run more eyecandy without any slowdowns.

With the F@H clients, overclocking will obviously help you do calculations faster so we can complete work units (WU in short) fast so they can issue new WU earlier. This will also help you increase the PPD (Points per day) output if you love to follow your stats and tweak them to improve the rate of climbing. However, there is also a dark side when you are overclocking past the actual components limits (due to heat or core limitations), there will be a big increase in risks of having issues delaying the return of WU and killing the PPD :

-Getting a EUE (Early Unit End)
-A bad crash corrupting the WU data causing it to restart from scratch or losing the WU totally, delaying the reissue of the same WU to another folder.
-Being unable to send completed WU or retrieve new WU.
-Getting errors like 0x7b, 0xFFFFFFFF, 0x1, 0x0 or other core errors causing it to delete the WU.
-Having issues starting or stopping mpiexec.exe or the FahCore_XX.exe threads.
-Other issues who cannot be solved by doing the basic troubleshooting steps.

In order to avoid those issues and be able to complete WU faster and reliably, we need to change our mindset how we consider a overclocking stable. This is important because the usual tools and things ppl in the overclocking work suggest to test for stability isn't good enough for a reliable folding. While it's not proved 100% accurate (beside banning any OC attempt and run it at stock speeds), those tips will help make the folding experience more enjoyable without any hair-pulling issues while ramping more points :

-Orthos, Prime95, OCCT, Super PI and other standard stress-test software is good to find out how stable your computer is running and to find the upper limits of your overclocking. However, those software isn't able to stress the computer the way folding is stressing so it isn't unexpected to see a computer stable with the above software for weeks fail at doing a single WU due to the various errors... The best testing we can do for a computer assigned to folding duty is to either run the F@H client itself and back down any overclocking till you can fold at least 2-3 WU in a row without any crashes or errors, or download StressCPU 2.0 which use the same Gromacs core as the F@H client. StressCPU will be able to stress it enough to fail a OC where others consider it stable. It will also test the memory hard enough to show any memory overclocking instabilities.

-Heat is also a enemy of any IC so it's important to cool it pretty well. Gaming, encoding, Photoshop or any other computer-intensive tasks is usually done in short amounts of time (2-3 hours at a time usually) so it may not be long enough to cause issues. Folding will stress your computer from starting the client till you shut down, which may be 24/7 if you always leave the computer on like most dedicated folders do. Being able to run stable for a few hours is different from being able to run 24/7 reliably and that's why all server hardware is designed to be run reliably 24/7 unlike desktop computers. When you overclock, keep this in mind and back down the overclocking a bit till you can run at a good temperature (usually under 50-60C under full load).

-Finally, if you installed a new OS on the computer then do tweakings, it would be a good idea to reinstall it again after finding the sweet spots, especially if you get lots of crashes, lockups or reboots while tweaking. This is because any crash or lockups has a big change of corrupting system files and this may play a key factor for getting F@H troubles. Installing a fresh copy ensure the files is not corrupted and will not mess with the running of a folding client. Sometime, the most occult issues nobody is unable to pinpoint a cause is often resolved by a OS reinstall.

I hope this will help fellow folders understand the issues related to OC and be less greedy when it come to top speed in favor of returning the results more reliably. More WU returned = more new WU available for grabs = more science being done = possibly a new cure earlier. If you are being greedy and ignore the crashes and other problems, you are actually causing the science to progress much slower !!

For any comments or suggestions with this article, PM me or email me at xilikon at hotmail dot com

P.S. : English isn't my first language so there may be typos/syntax/grammar errors. Feel free to correct it and rephrase to make it more accurate

 

[Tigerbiten]

Nice article.

My sequence for testing an overclocked folding box is .......
Build box.
Find the max speed it successfuly POST's at.
Back it down to stock speeds and boot Memtest 86+ from either floppy or cd.
Will it run error free.
If it does then overclock CPU & memory and retest.
Play around with the settings untill you find the best combo for CPU & memory that will pass at least three passes errors free.
Once your happy the set it back to stock and install the OS.
Patch it, update drivers, etc, etc as needed.
Now overclock it back to the sweet spot and start retasting.
I now use Orthos as the fifth test you can pick is the Gromacs stress test and it fires up as many copies of that as you have cores.
If it runs 24 hour error free and the temps are good then your good to fold.

One thing I've noticed with the SMP client with an overclocked Windoze box.
If you put a shortcut to it in the run -> programs -> startup menu, then you can get an error every time you restart because the client can start before Windoze is ready for it.
Give the boxen a couple of minutes to finish loading drivers, sevices, etc and the client starts fine.
Luck ...........

 

[BillR]

Nice article. One thing I've noticed with the SMP client with an overclocked Windoze box. If you put a shortcut to it in the run -> programs -> startup menu, then you can get an error every time you restart because the client can start before Windoze is ready for it.Give the boxen a couple of minutes to finish loading drivers, sevices, etc and the client starts fine.
Luck ...........

Stealth mode^^^?

Aside from your other good advice I have found that many people have slow folders for just the reason you stated. In fact it’s why I never start it as a service.

That brings me to the much forgotten –forceasm flag. If you have to run folding as a service (which leaves out the SMP client) don’t forget to add that flag to your service configuration.

Each client of late auto starts with –forceasm on, but any problems the client sees it will shut that part down and you lose about half your folding speed. This is particularly important for laptop users or people who reboot frequently.

As for the –advmethods flag and SMP, I have yet to see it make a difference in which work units you get, at least for now.

Just some quick thoughts to add the good info above

 

[Imitation]

I would also like to add that anyone that runs hamachi might have issues with the winSMP client. I've found that whenever I've got hamachi running but not logged on, I'm ok, but logging onto hamachi essentially changes the IP address, hosing your folding processes. So what I do now is just stay logged into hamachi at all times since I created a new network recently with only people I trust. I haven't had any core stopping issues for the last month or so. I tried EVERYTHING before I figured this out, slowing up clocks, more voltage for more stability, yet logging in and out of hamachi was my undoing. This isn't really overclocking related, but related to overall folding stability

 

[Xilikon]

I would also like to add that anyone that runs hamachi might have issues with the winSMP client. I've found that whenever I've got hamachi running but not logged on, I'm ok, but logging onto hamachi essentially changes the IP address, hosing your folding processes. So what I do now is just stay logged into hamachi at all times since I created a new network recently with only people I trust. I haven't had any core stopping issues for the last month or so. I tried EVERYTHING before I figured this out, slowing up clocks, more voltage for more stability, yet logging in and out of hamachi was my undoing. This isn't really overclocking related, but related to overall folding stability

that's a good point, especially with the SMP client who rely on the networking stack for the communication between cores. I should also add that in some cases, overclocking could cause networking issues, especially on those motherboard who have no pci/agp lock. I cannot elaborate on this since it's pretty rare but should be a thing to consider when troubleshooting folding issues.

 

[Tigerbiten]

One last point on the SMP client ......
It must setup to accept big packet work units.
The downloading of work does not go over the small packets limits but the uploading of finished work units can.
So with the config switch set the wrong way it will hang trying to upload work.

On to the GPU client .............

The client needs Dx9 to poll the CPU.
This will load one CPU core at 100% for each client at the moment.
Dx9 is only single threaded so running two GPU clients on two vid cards gives you a minimum slowdown of ~10% per client.
The more the vidio cards differ in speed the worse the slow down.

Altering the screen res or logging in/out causes the GPU client to shut down.
So you cannot use remote desktop to check your Windozes boxen are ok if running a GPU client.
Also if a game alter the screen res then you will need to restart the GPU client after you have finished playing.

I found that if your vid card was overclocked to the limit the the GPU client would run fine untill it glitched.
Once it glitched then it would churn out EUE untill restarted.
Once it was restarted then it would run fine untill the next glitch.

As for running a GPU client with an SMP client, it possible.
BUT its no longer a set and forget setup.
The affinity GPU client needs to be set so it only uses one CPU core.
Then the affinity of the SMP client needs to be set so that it has any remaining cores.
The GPU client can be locked to one core once set, the SMP client cannot so needs to be reset for every protien.
Depending on your CPU & vid card this type of setup could give you an extra 200-300 PpD.
But only run this setup if both clients can return work before their first deadlines otherwise its bad for the science as that work will get reasigned.

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