P-1 factoring anyone? - Page 100

kladner · 2012-01-31, 00:55

Quote:

Originally Posted by Mini-Geek

Here's a possibility that affected me recently: if Prime95's RollingAverage (a measure of how fast you really are vs how fast it expects you to be) is off, it will choose bounds differently. I saw Prime95 choose much higher bounds when I manually bumped up the rolling average to be more accurate (it was about half what it should've been).

Hmm. This makes me wonder. A while back, I corrected the ~10GHz which P95 was reporting for a PhenomII running at 3510MHz. Now this is what local.txt shows:

Code:

CpuSpeed=3512
OldCpuSpeed=3512
NewCpuSpeedCount=0
NewCpuSpeed=0
RollingAverage=953

Is RollingAverage supposed to be showing something remotely related to the actual core speed?

Dubslow · 2012-01-31, 01:00

Yes, I think your example proves that. Some sort of override would be nice.

chalsall · 2012-01-31, 02:06

Quote:

Originally Posted by kladner

Is RollingAverage supposed to be showing something remotely related to the actual core speed?

My inference is that this is a value which is suppost to represent the precentage of the CPU actually available, times 10.

It seems that Prime95/mprime sometimes makes mistakes on this (sometimes for very short periods of time) and then makes important decisions based on this value.

Prime95 · 2012-01-31, 03:36

Quote:

Originally Posted by kladner

Is RollingAverage supposed to be showing something remotely related to the actual core speed?

No. It is a measure of actual throughput versus expected throughput. Your rolling average of 953 means prime95 is getting 95.3% of the expected throughput.

The rolling average is used to compute expected completion dates (and thus deciding whether to unreserve assignments). The server uses the value to decide what kind of work assignment "makes the most sense".

Rolling average is not used to compute P-1 bounds.

kladner · 2012-01-31, 03:42

Thanks for the clarification.

Dubslow · 2012-01-31, 04:27

I guess that if the core speed is way off, then the RollingAverage will be way off indirectly. (If your core speed was 10GHz, then I'm sure your RA was somewhere around ~300.)

kladner · 2012-01-31, 04:32

Quote:

Originally Posted by Dubslow

I guess that if the core speed is way off, then the RollingAverage will be way off indirectly. (If your core speed was 10GHz, then I'm sure your RA was somewhere around ~300.)

True a month ago. But 95.3% is probably pretty accurate when one considers restarts and demanding software (Photoshop, Camera Raw, etc) taking a chunk out of ideal performance.

cheesehead · 2012-01-31, 07:00

Quote:

Originally Posted by bcp19

I was just noticing that my P-1 machine has had a weird bounds with no change to the memory, can anyone explain if this is something other than how far the exp was TF'd?

Code:

[Sat Jan 28 09:29:51 2012]
P-1 found a factor in stage #2, B1=430000, B2=8707500.
UID: bcp19/HP-NEW, M45048023 has a factor: 360751991413212824008821379007
[Sat Jan 28 17:23:52 2012]
UID: bcp19/HP-NEW, M45122951 completed P-1, B1=340000, B2=5780000, E=6, We4: 8F2BB36D
[Sun Jan 29 04:22:56 2012]
UID: bcp19/HP-NEW, M45158209 completed P-1, B1=430000, B2=8707500, E=12, We4: 90ECBFD6
[Sun Jan 29 15:21:47 2012]
UID: bcp19/HP-NEW, M45159679 completed P-1, B1=430000, B2=8707500, E=12, We4: 90C4BFFD
[Mon Jan 30 02:20:51 2012]
UID: bcp19/HP-NEW, M45163583 completed P-1, B1=430000, B2=8707500, E=12, We4: 90E7BF9E

This looks exactly like a situation in which M45122951 has been TFed to a higher level than the other exponents have been TFed.

Upon checking the status report, I find that M45122951 has been TFed to 2^75, while M45158209, M45159679, and M45163583 have each been TFed to only 2^72. M45048023 had almost certainly also been TFed to only 2^72 before this P-1 run.

Higher TF already done means that the probability of finding a factor with a particular set of P-1 bounds is lower than it would have been if less TF had already been done. Thus, the optimum B1/B2 combination of best balance for the high-TFed exponent occurs at lower bounds than for the lower-TFed exponents.

James Heinrich · 2012-01-31, 12:50

Quote:

Originally Posted by cheesehead

Quote:

Originally Posted by bcp19

...can anyone explain if this is something other than how far the exp was TF'd?

Upon checking the status report, I find that M45122951 has been TFed to 2^75, while M45158209, M45159679, and M45163583 have each been TFed to only 2^72. M45048023 had almost certainly also been TFed to only 2^72 before this P-1 run.

Thanks for looking that up. I didn't bother checking TF levels since bcp19 seemed to imply that he'd ruled out the obvious explanation.

Prime95 · 2012-01-31, 15:25

Quote:

Originally Posted by kladner

True a month ago. But 95.3% is probably pretty accurate when one considers restarts and demanding software (Photoshop, Camera Raw, etc) taking a chunk out of ideal performance.

The biggest source of deviation from 1000 is that prime95's initial estimate is not very accurate.

bcp19 · 2012-01-31, 15:35

Quote:

Originally Posted by James Heinrich

Thanks for looking that up. I didn't bother checking TF levels since bcp19 seemed to imply that he'd ruled out the obvious explanation.

Sorry, I hadn't ruled that out, it just seemed an awful big jump in bounds for just a few extra levels.

2012-01-31, 01:00	#1091
Dubslow Basketry That Evening! "Bunslow the Bold" Jun 2011 40<A<43 -89<O<-88 3·29·83 Posts	Yes, I think your example proves that. Some sort of override would be nice.

2012-01-31, 03:42	#1094
kladner "Kieren" Jul 2011 In My Own Galaxy! 2·3·1,693 Posts	Thanks for the clarification.

2012-01-31, 04:27	#1095
Dubslow Basketry That Evening! "Bunslow the Bold" Jun 2011 40<A<43 -89<O<-88 3×29×83 Posts	I guess that if the core speed is way off, then the RollingAverage will be way off indirectly. (If your core speed was 10GHz, then I'm sure your RA was somewhere around ~300.)