-forcesse and AMD CPUs

’

’m‚³‚ñ

Supreme [H]ardness
Joined
Jun 20, 2003
Messages
7,300
Okay, I'm not quite understanding what happens here... When you are using -forcesse on a CPU and it 'locks up', does the entire system go down, or does F@H just stop working? I've had both of my AMD machines with F@H4 completely die, and the only thing changed was the client version and the flag.

Cliffs notes: Upgraded two AMD systems to F@H 4
Both machines completely died.
Is this the -forcesse flag or something else?
 
When an AMD based machine with Tbred "B" cores or beyond using the SSE flag locks up the entire box freezes requiring a restart. In every case I've seen the display is still there showing the desktop, but the box is hard locked. The issue doesn't really lie with any particular client, but the actual core, ( FAHcore_78.exe which is GROMACS) program. So I may still be running 3.25 and you could be using 4.0 in the client but we both would be running the same core program. The jury is still out where the problem is coming from, be it AMD's SSE in theses newer cores or something in the Stanford code. Although Stanford's contention is if it was in their code then why doesn't the problem also crop up in Palomino AMD cores that also use SSE as well.
 
I have 3 xp 2400 machines on various chipsets that hard lock. I also have an amd 64 3000 that also hard locks. I have 3 xp 2500 and 1 xp 2800 that have never locked up. I cannot figure it out for the life of me. They run for weeks if I disable the SSE flag. With SSE flag enabled they always crash within a day or two. Needless to say my Pentiums have never crashed.
 
I've got over a dozen machines folding (Celerons, P4's, Durons, and PIII's, all with -forecasm or -forceSSE,
and the only machine that has ever locked up is my AMD 2600.
The lockups are very intermittent, however, perhaps once per week.
When I reboot, it finishes the protein so I've kept SSE running.

90000.gif
 
Originally posted by Celerator
I've got over a dozen machines folding (Celerons, P4's, Durons, and PIII's, all with -forecasm or -forceSSE,
and the only machine that has ever locked up is my AMD 2600.
The lockups are very intermittent, however, perhaps once per week.
When I reboot, it finishes the protein so I've kept SSE running.

90000.gif
Click to expand...

Ah, okay. Thanks. I guess it wasn't F@H then. =\ My main box is a Barton 2500+, it was completely off as if it took a powerloss, and the second box is a Palomino, so it's not really going to be effected by this. The dell PIII is still chugging away... :p
 
If you have machines that are experiencing the problem please remember to post their hardware data in this thread at the Stanford Folding forum. They are trying to figure out why it's occuring.

http://forum.folding-community.org/viewtopic.php?t=6682

Other general comments and discussion about the issue are taking place in this thread

http://forum.folding-community.org/viewtopic.php?t=6683



I noticed this post in the first thread that was dated 15 Jan.

AMD thanks the Folding at Home users for providing information about the freezing problem. Thanks to Prof Pande, AMD has reproduced the problem in our Austin labs. Although not an official workaround, we have observed that if the console application is launched without the -forceasm switch, or if the GUI version advanced properties setting to enable advanced optimizations is not enabled, then the freeze does not occur.

We apologize for any inconvenience and will be working with Prof. Pande as soon as a better workaround is available.

AMD_Mike
Click to expand...
 
On my amd 2000+xp when -advmethods and -forcesse is enabled and i've Oced it it will lock up every time but when running all stock speed it actually folds faster then when i would OC it so i've left it now back at stock speed and it's been running fine for about a month now with no problems
 
On a 1 Ghz Morgan core Duron I've been experiencing many, sometime 1 a day, lockups, and just got my second :
Folding@home Core Shutdown: EARLY_UNIT_END
Click to expand...
since I've moved to v4 -
I just moved back to v3.25.
 
Originally posted by CIWS
If you have machines that are experiencing the problem please remember to post their hardware data in this thread at the Stanford Folding forum. They are trying to figure out why it's occuring.

http://forum.folding-community.org/viewtopic.php?t=6682

Other general comments and discussion about the issue are taking place in this thread

http://forum.folding-community.org/viewtopic.php?t=6683



I noticed this post in the first thread that was dated 15 Jan.
Click to expand...

I can verify, both my new nf7-s and an7 systems (with 2500+ barton's) would hard lock with both -forceasm and -forcesse. Tried just forceasm, system would not lock but was also "slower." Then tried just -forcesse all by itself. Seem to get the speed boost and I have been lock free for 7 days now!!!

So the question is, what does the -forceasm get me???? I see the speed boost with just -forcesse, so I guess we just don't care about -forceasm????
 
Originally posted by gnewbury
On a 1 Ghz Morgan core Duron I've been experiencing many, sometime 1 a day, lockups, and just got my second :
since I've moved to v4 -
I just moved back to v3.25.
Click to expand...

Interesting. My Palomino did that once, a while ago. First time after I upgraded to F@H 4 and turned the sse boost on. Hasn't done that since. Then agin, since then, I haven't gotten ANY work done because after every WU it finishes, my firewall refuses to let it send the data to the server because 'OMG DIFFERENT VERSION DO YOU WANT TO LET IT USE THE INTERNET BECAUSE IT"S CHANGED AND STUFF AND THINGS AND OMG'... Okay, time to eat, the little things are driving me nuts again.
 
Originally posted by stnelson9
I can verify, both my new nf7-s and an7 systems (with 2500+ barton's) would hard lock with both -forceasm and -forcesse. Tried just forceasm, system would not lock but was also "slower." Then tried just -forcesse all by itself. Seem to get the speed boost and I have been lock free for 7 days now!!!

So the question is, what does the -forceasm get me???? I see the speed boost with just -forcesse, so I guess we just don't care about -forceasm????
Click to expand...

-forceasm means use the default optimization when running GROMACS core, regardless of prior system lockups or crashes. On a modern AMD processor that is 3DNow! instruction set. On an Intel processor that is SSE instruction set.

-forcesse means use SSE instruction set, regardless of prior crashes. It takes precedence if used in conjunction with -forceasm.

The SSE code in the GROMACS core runs faster than the 3DNow! code. I don't know if that says anything about the relative performance of SSE versus 3DNow, but since the optimizations are developed for SSE it comes as no surprise that it is faster. What you are seeing with -forceasm is the core defaulting to using the 3DNow! instruction set. It runs slower but shows no problems related to the lockups of AMD processors running SSE optimizations.
 
Originally posted by Chugiak
-forceasm means use the default optimization when running GROMACS core, regardless of prior system lockups or crashes. On a modern AMD processor that is 3DNow! instruction set. On an Intel processor that is SSE instruction set.

-forcesse means use SSE instruction set, regardless of prior crashes. It takes precedence if used in conjunction with -forceasm.

The SSE code in the GROMACS core runs faster than the 3DNow! code. I don't know if that says anything about the relative performance of SSE versus 3DNow, but since the optimizations are developed for SSE it comes as no surprise that it is faster. What you are seeing with -forceasm is the core defaulting to using the 3DNow! instruction set. It runs slower but shows no problems related to the lockups of AMD processors running SSE optimizations.
Click to expand...

So to summarize, am I correct in saying that: -forcesse -forceasm on the command line is just: -forcesse????

If that is the case, then I don't see why I am not seeing the lockups with just the -forcesse flag on. Unless there is a specific bug tied to having -forceasm on the command line, which I doubt... I would have to say it is coincidence then, that my lockups have stopped by removing the -forceasm flag...
 
Originally posted by stnelson9
So to summarize, am I correct in saying that: -forcesse -forceasm on the command line is just: -forcesse????

If that is the case, then I don't see why I am not seeing the lockups with just the -forcesse flag on. Unless there is a specific bug tied to having -forceasm on the command line, which I doubt... I would have to say it is coincidence then, that my lockups have stopped by removing the -forceasm flag...
Click to expand...

I had two instances of F@H 4 funning after I updated and forgot to remove the second entry from the Start folder. One ran with -forcesse and the other with -forceasm, and both would hardlock as soon as they loaded. Now I'm just running one with -forcesse and it works just fine on my 2500+ @ 2.3 ghz. Hasn't hard locked yet, except for the mysterious time it shut down...

My PIII is still chugging along. I'll check it's uptime later today. :)
 
Originally posted by stnelson9
So to summarize, am I correct in saying that: -forcesse -forceasm on the command line is just: -forcesse????
<snip>.
Click to expand...

I assume so. At least that is how I would write the code. But I really don't know.
 
From sources at folding community.

-forceSSE
On machines with an AMD processor, Core_78 gives priority to 3DNow over SSE -- this overrides that
After a crash, -forceasm will re-enable the default assembly optimizations, which is 3DNow for AMD or SSE for Intel.

From MannyMax
4.00
Are assembly optimizations disabled (due to crashing, configuration, or availability)?
If yes, use x87.

If no, is this an Intel CPU?
If yes, use SSE.

If no, is this a CPU that supports SSE?
If no, use 3DNow!+.

If yes, is "-forceSSE" being used?
If yes, use SSE.
If not, use 3DNow!+.

I hope you can follow the above flowcharts... they may be a bit confusing, but they're the best I can do for now.
Click to expand...
 
You must log in or register to reply here.
Back
Top