KWSN Orbiting Fortress

[Putting_things_on_top]

I'm not sure if this will work for everyone, but it really has helped me maintain a more consistent daily output.
I have been plagued by NAN/EUE errors (F@H) on GPU folding...and after 5 or 6 successive EUE's, the client suspends itself for 24 hours.
This really cranked me off for the past few months!!!

But NOW, I have found a solution that works (at least for me!)

• I added the -oneunit startup switch/flag (amongst others) to my GPU client configs
  
• I set-up an infinite schedule to start each GPU client every 5 minutes (unless already running) thru Windows Task Scheduler
  NOTE: You must start GPU clients thru Task Scheduler only in order for it to manage successive restarts correctly

This set-up has some advantages:

1. Every WU completion -or- error causes the client to stop cleanly (no EUE counter)
   
2. Each GPU can get up to 5 minutes to cool down between WU's
   
3. I get to sleep better at night without feeling compelled to restart failed clients in the middle of the night

Since I've set this up (almost two weeks ago), I have been getting around 20,000+ points per day (average).
Before, I think I was averaging 13,000 points (and with a lot of manual intervention on my part).
Yes, I still get NAN errors, but they only stop a GPU for - at most - 5 minutes!!!
Hope this helps others
_________________
Click here for...KWSN F@H team summary at EOC

Or here for...KWSN F@H team overtake at EOC

[Idan]

That's a really good tip!

Thanks a bunch!
_________________
Anyone for Crunch?

[Putting_things_on_top]

What IS a NAN (you might ask)...

It really means "Not A Number", and (at least in terms of GPU folding) is the result of a double-precision floating-point calculation whose results are either too large or too small to be handled with any reasonable accuracy by the hardware.

This is why we almost always see a "UNSTABLE_MACHINE" return code in the logfile. And when F@H refers to a "machine", it doesn't necessarily mean that the hardware is malfunctioning. Some of these folding applications are using some kind of quasi finite-state-machine-theory style of programming (too complicated to explain here)...so they are really talking about the virtual machine that was created within the app.

Almost all Nvidia GPUs are incapable of handling double-precision operations to the requisite degree of accuracy; a few of the ATI cards are designed to natively process double-precision, but these are a limited set of models (not necessarily the high-end ones, either). As for Nvidia, they claim to have DPFPO architected into their (soon-to-be-released) Fermi series.

Additionally, some of the F@H projects are notoriously dependent on DPFPO (the 576x project series, for example), and it gets to be a crap-shoot as to whether or not a WU will succeed on a GPU. I have had WUs go to 85% and then crap-out...and I don't even get partial credit!

We generally do not see NAN errors returned from CPU clients because all 'modern' CPUs already have DPFPO built-in (much older CPUs relied on a separate "math-coprocessor" chip to achieve this functionality).

Oh, DPFPO = double precision floating point operations
_________________
Click here for...KWSN F@H team summary at EOC

Or here for...KWSN F@H team overtake at EOC