> Something peculiar I have noticed with my Intel P4's is over at the Seti Site
> the 1 Prescott P4 3.4 I have will run the WU's 30-60 min's faster than the 3
> Northwood P4 3.4's I have will. But here at the Einstein Site the Northwoods
> will out run the Prescott by 30-60 min's .... Weird to say the least ... :)
I'm not a PC expert by any means, but didn't the Northwoods have a shorter pipeline then the Prescotts?
Whatever it is, it has something to do with the type of calculations performed. What I think I am seeing is that the faster machines with the longer pipelines are having to dump more data much more often because of cache misses or pipeline stalls. The Prescott has a 30 stage pipeline instead of (I think) 20 for the the Northwood. Whenever there is a cache miss or pipeline stall, the Prescott may lose up to 30 cycles while the Northwood loses up to 20 cycles.
Apparently the type of calculations involved for current calculations, may result in more misses for predicting what type of information the CPU will need next. Whenever there is a miss on the prediction, then the pipeline must be cleared of the incorrect intructions/data so that the correct instructions/data can be loaded. The bigger the pipeline, the more CPU cycles are wasted waiting for the intructions to be reloaded. - The shorter the pipeline, the fewer the wasted cycles. - - That is pretty simplistic, and the article will explain it better. Essentially I am seeing that CPU's with shorter pipelines such as AMD's are overall more effecient on Einstein, despite slower CPU speeds. For calculations where the data needed by the CPU can be more accurately predicted (less random) the Intel chips excell. For more random instructions and data feeds, the AMD's excell.
At least I think that is the general idea.
Anyone running a 1.42ghz G4? I am curious how it compares to the 1.6 G5 on Einstein. The G4 has a shorter pipeline. Einstein calculations don't rely heavily on memory or bus speed, so I am wondering if the G4 would be near the speed of the 1.6 G5.
I run about 4 to 5 hrs per WU with my AMD 64 3200+. Weird thing is though, i've only completed 3 of my 6 WU's. Those other 3 exited with an incorrect function. (0x1)
> I run about 4 to 5 hrs per WU with my AMD 64 3200+. Weird thing is though,
> i've only completed 3 of my 6 WU's. Those other 3 exited with an incorrect
> function. (0x1)
I just had a look at your failed workunits. You can do this too. Look at the that you can find within the 'results' section of the webpage.
A typical failed job is shown below. The job ran out of memory precisely because of the problem described on the Einstein@Home front page, which Xavier Siemens is currently fixing.
One of the interesting features of our work is that the methods that we use to analyze the data depend upon the features of the data. But until we have done the analysis, we often don't know what all of these salient features are. So these kinds of issues come as no surprise to us: our analysis is driven by the data and often needs tuning, fixing and refinement as we understand these features of the data.
Thank you again for helping us with the testing: I am optimistic that we'll have these problems under control soon.
Resuming computation at 11629/1266478/1267294
Resuming computation at 21319/2260539/2438681
Resuming computation at 31006/3201597/3253402
Resuming computation at 1459/272639/273277
Resuming computation at 31006/3201597/3253402
Resuming computation at 12688/1418556/1419740
Resuming computation at 31006/3201597/3253402
Resuming computation at 22907/2428058/2428058
Error: ran out of memory ... goodbye.
Glad that my "failed" WU's are helping out, because after erroring out on 4 of the 7 total WU's so far, one starts to think on whether or not one is making a contribution. {edit} All of my failed WU's are ran out of memory errors, even though some have completed those WU's. {/edit}
@Shaktai: that's pretty right about Northwood vs. Presshot. Also, Presshot is having larger, but slower cache.
=========
Your right about the Cache Honza & it's Prescott not Presshot ... ;)
But at the Seti Site like I said it seems to be an advantage running the type of WU's they do there but here at Einstein it's seems to be a drawback ...
> But at the Seti Site like I said it seems to be an advantage running the type
> of WU's they do there but here at Einstein it's seems to be a drawback ...
That is pretty much it. The einstein calculations favor AMD like architecture and older (shorter pipeline) Intell chips, more then the new Inel chips. It will vary with other projects. I agree the Intel's do quite well at SETI and Climate.
Well I guess I'll have to slack along at 2 WU's every 10 to 10 1/2 hours with that Computer, oh wait, that beats almost 100% of the CPU's out there anyway ... hehehe "Evil Grin" ... ;)
> Something peculiar I have
)
> Something peculiar I have noticed with my Intel P4's is over at the Seti Site
> the 1 Prescott P4 3.4 I have will run the WU's 30-60 min's faster than the 3
> Northwood P4 3.4's I have will. But here at the Einstein Site the Northwoods
> will out run the Prescott by 30-60 min's .... Weird to say the least ... :)
I'm not a PC expert by any means, but didn't the Northwoods have a shorter pipeline then the Prescotts?
Whatever it is, it has something to do with the type of calculations performed. What I think I am seeing is that the faster machines with the longer pipelines are having to dump more data much more often because of cache misses or pipeline stalls. The Prescott has a 30 stage pipeline instead of (I think) 20 for the the Northwood. Whenever there is a cache miss or pipeline stall, the Prescott may lose up to 30 cycles while the Northwood loses up to 20 cycles.
Here is an article that kind of explains it: "http://www.pcmech.com/show/processors/715/"
Apparently the type of calculations involved for current calculations, may result in more misses for predicting what type of information the CPU will need next. Whenever there is a miss on the prediction, then the pipeline must be cleared of the incorrect intructions/data so that the correct instructions/data can be loaded. The bigger the pipeline, the more CPU cycles are wasted waiting for the intructions to be reloaded. - The shorter the pipeline, the fewer the wasted cycles. - - That is pretty simplistic, and the article will explain it better. Essentially I am seeing that CPU's with shorter pipelines such as AMD's are overall more effecient on Einstein, despite slower CPU speeds. For calculations where the data needed by the CPU can be more accurately predicted (less random) the Intel chips excell. For more random instructions and data feeds, the AMD's excell.
At least I think that is the general idea.
Anyone running a 1.42ghz G4? I am curious how it compares to the 1.6 G5 on Einstein. The G4 has a shorter pipeline. Einstein calculations don't rely heavily on memory or bus speed, so I am wondering if the G4 would be near the speed of the 1.6 G5.
Team MacNN - The best Macintosh team ever.
I run about 4 to 5 hrs per WU
)
I run about 4 to 5 hrs per WU with my AMD 64 3200+. Weird thing is though, i've only completed 3 of my 6 WU's. Those other 3 exited with an incorrect function. (0x1)
@Shaktai: that's pretty right
)
@Shaktai: that's pretty right about Northwood vs. Presshot. Also, Presshot is having larger, but slower cache.
My general experience is that P4 are doing better on CPDN and SETI; AMD better on Einstein and ProteinPredictor.
> I run about 4 to 5 hrs per
)
> I run about 4 to 5 hrs per WU with my AMD 64 3200+. Weird thing is though,
> i've only completed 3 of my 6 WU's. Those other 3 exited with an incorrect
> function. (0x1)
I just had a look at your failed workunits. You can do this too. Look at the that you can find within the 'results' section of the webpage.
A typical failed job is shown below. The job ran out of memory precisely because of the problem described on the Einstein@Home front page, which Xavier Siemens is currently fixing.
One of the interesting features of our work is that the methods that we use to analyze the data depend upon the features of the data. But until we have done the analysis, we often don't know what all of these salient features are. So these kinds of issues come as no surprise to us: our analysis is driven by the data and often needs tuning, fixing and refinement as we understand these features of the data.
Thank you again for helping us with the testing: I am optimistic that we'll have these problems under control soon.
Cheers,
Bruce
4.16
Incorrect function. (0x1) - exit code 1 (0x1)
1
0
Resuming computation at 11629/1266478/1267294
Resuming computation at 21319/2260539/2438681
Resuming computation at 31006/3201597/3253402
Resuming computation at 1459/272639/273277
Resuming computation at 31006/3201597/3253402
Resuming computation at 12688/1418556/1419740
Resuming computation at 31006/3201597/3253402
Resuming computation at 22907/2428058/2428058
Error: ran out of memory ... goodbye.
Director, Einstein@Home
Glad that my "failed" WU's
)
Glad that my "failed" WU's are helping out, because after erroring out on 4 of the 7 total WU's so far, one starts to think on whether or not one is making a contribution. {edit} All of my failed WU's are ran out of memory errors, even though some have completed those WU's. {/edit}
@Shaktai: that's pretty right
)
@Shaktai: that's pretty right about Northwood vs. Presshot. Also, Presshot is having larger, but slower cache.
=========
Your right about the Cache Honza & it's Prescott not Presshot ... ;)
But at the Seti Site like I said it seems to be an advantage running the type of WU's they do there but here at Einstein it's seems to be a drawback ...
I like 'Presshot'...the term
)
I like 'Presshot'...the term is much more accurate. Once TDP goes down, i will recognize it as Prescott again :-)
> Your right about the Cache Honza & it's Prescott not Presshot ... ;)
Double post deleted, sorry...
)
Double post deleted, sorry...
> But at the Seti Site like I
)
> But at the Seti Site like I said it seems to be an advantage running the type
> of WU's they do there but here at Einstein it's seems to be a drawback ...
That is pretty much it. The einstein calculations favor AMD like architecture and older (shorter pipeline) Intell chips, more then the new Inel chips. It will vary with other projects. I agree the Intel's do quite well at SETI and Climate.
Team MacNN - The best Macintosh team ever.
Well I guess I'll have to
)
Well I guess I'll have to slack along at 2 WU's every 10 to 10 1/2 hours with that Computer, oh wait, that beats almost 100% of the CPU's out there anyway ... hehehe "Evil Grin" ... ;)