Will Any Current 100M Digit LL Tests Finish?

[jinydu]

Originally Posted by gd_barnes

Am I missing something here?

Well the question did say "current". As I mentioned in a later post, I'm only asking about tests registered on or before 2008. So there isn't much time left to future speed increases.

Still, you're right that the estimated time to complete a 100M digit test is still well within a human lifetime. I think it's mainly a question about human patience, although another consideration is the chance that the computer could fail before the test finishes. I don't know whether any of the users who have registered such a test understand how long it will take, and are sincerely intending to carry through with it anyway. For instance...

Originally Posted by James Heinrich

As it relates to this thread, I will not be finishing my assigned LL test, I'm just running through TF and P-1 (ETA: end of Jan 2009).

As for your question about a billion digit test... So far that isn't even possible on Prime95 because it only supports test up to M999,999,999.

[James Heinrich]

Will one of the current ~480 assigned tests be finished by the user who currently has it reserved? I say very probably not. Will a 100M test be finished in the next 5 years or so? Almost definitely. 5 years from now the per-core processing power will likely be ~10x the Core2 level; moreover 32-core CPUs will almost certainly be around by then. Assuming those two predictions are true then the current 5+ year test will turn into a 1-week test on then-current hardware.

[petrw1]

Originally Posted by petrw1

Using Time... on a:

Model Intel(R) Pentium(R) 4 CPU 3.40GHz 
Features Single core, 2 hyperthreaded CPUs, Prefetch,SSE,SSE2 
Speed 3.391 GHz (3.533 GHz P4 effective equivalent) 
L1/L2 Cache 16 / 1024 KB 
Computer Memory 1016 MB   configured usage 200 MB day / 200 MB night

I guess at the exponent to be 100M. I might be a little off.

[Nov 28 15:34] Setting affinity to run helper thread 1 on logical CPUs 0,1
[Nov 28 15:34] Using FFT length 20M, 2 threads
[Nov 28 15:35] p: 332190991.  Time: 885.703 ms.
[Nov 28 15:35] p: 332190991.  Time: 879.040 ms.
[Nov 28 15:35] p: 332190991.  Time: 877.096 ms.
[Nov 28 15:35] p: 332190991.  Time: 884.255 ms.
[Nov 28 15:35] p: 332190991.  Time: 885.846 ms.
[Nov 28 15:35] p: 332190991.  Time: 874.900 ms.
[Nov 28 15:35] p: 332190991.  Time: 877.598 ms.
[Nov 28 15:35] p: 332190991.  Time: 884.331 ms.
[Nov 28 15:35] p: 332190991.  Time: 874.370 ms.
[Nov 28 15:35] p: 332190991.  Time: 877.754 ms.
[Nov 28 15:35] Iterations: 10.  Total time: 8.801 sec.
[Nov 28 15:35] Estimated time to complete this exponent: 3383 days, 18 hours, 28 minutes.

Seems like a rather large FFT=20M.
Over 9 years.

P.S. I did my first 10M on a pIII 400Mhz and it only took 15 months.

On my Q9550 use 1-4 cores respectively the times for 100M are:
2072, 1071, 856, 687 Days.
So running all 4 cores would have me theoretically done in just under 2 years.

[James Heinrich]

Originally Posted by petrw1

On my Q9550 use 1-4 cores respectively the times for 100M are:
2072, 1071, 856, 687 Days.
So running all 4 cores would have me theoretically done in just under 2 years.

On my Q6600 @ 3.5GHz the numbers come up as:

Total time: 4.470; 2.426; 2.067; 1.741
ETC: 1718 days; 932 days; 794 days; 669 days

for 1 through 4 cores (my 4-core performance is relatively limited by slow RAM). But the estimated date of completion of M332203901 is 16-Apr-2013 right now (if I only use 1 core the estimate is sometime in 2020). I'm not sure on why the widely different estimates (2 years vs 5 years)...?

[davieddy]

Originally Posted by gd_barnes

Am I missing something here? Is there some reason to think that the testing time will vary more than by the square of the exponent?

Gary

The time per iteration goes more than proportionally
to the exponent

However I read in Knuth that for some reason you could
multiply as n rather than nlogn but I didn't see why.

[cheesehead]

Originally Posted by davieddy

However I read in Knuth that for some reason you could multiply as n rather than nlogn but I didn't see why.

Multiplication can approach O(n) by using more complicated methods than we are using in GIMPS (or by dedicating more and more gates on the CPU chip to multiplication). Knuth explains those approaches later in the chapter ("4.3.3 E. Multiplication in real time"). "4.3.3 C. Discrete Fourier transforms" covers what GIMPS uses.

[henryzz]

[Main thread Nov 29 10:43] Starting worker.
[Work thread Nov 29 10:43] Worker starting
[Work thread Nov 29 10:43] Setting affinity to run worker on logical CPU #0
[Work thread Nov 29 10:43] Using FFT length 20M
[Work thread Nov 29 10:43] p: 332190991.  Time: 521.265 ms.
[Work thread Nov 29 10:43] p: 332190991.  Time: 524.189 ms.
[Work thread Nov 29 10:43] p: 332190991.  Time: 521.293 ms.
[Work thread Nov 29 10:43] p: 332190991.  Time: 520.986 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 520.824 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 521.945 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 520.860 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 520.538 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 537.287 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 526.447 ms.
[Work thread Nov 29 10:44] Iterations: 10.  Total time: 5.236 sec.
[Work thread Nov 29 10:44] Estimated time to complete this exponent: 2012 days, 23 hours, 57 minutes.
[Work thread Nov 29 10:44] Setting affinity to run helper thread 1 on logical CPU #1
[Work thread Nov 29 10:44] Using FFT length 20M, 2 threads
[Work thread Nov 29 10:44] p: 332190991.  Time: 287.737 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 281.572 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 282.587 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 281.403 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 283.385 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 286.273 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 282.167 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 284.794 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 283.142 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 284.227 ms.
[Work thread Nov 29 10:44] Iterations: 10.  Total time: 2.837 sec.
[Work thread Nov 29 10:44] Estimated time to complete this exponent: 1090 days, 21 hours, 9 minutes.
[Work thread Nov 29 10:44] Setting affinity to run helper thread 1 on logical CPU #1
[Work thread Nov 29 10:44] Setting affinity to run helper thread 2 on logical CPU #2
[Work thread Nov 29 10:44] Using FFT length 20M, 3 threads
[Work thread Nov 29 10:44] p: 332190991.  Time: 273.486 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 273.899 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 282.173 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 276.716 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 303.454 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 278.922 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 251.117 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 269.764 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 257.384 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 255.954 ms.
[Work thread Nov 29 10:44] Iterations: 10.  Total time: 2.723 sec.
[Work thread Nov 29 10:44] Estimated time to complete this exponent: 1046 days, 21 hours, 20 minutes.
[Work thread Nov 29 10:44] Setting affinity to run helper thread 1 on logical CPU #1
[Work thread Nov 29 10:44] Setting affinity to run helper thread 2 on logical CPU #2
[Work thread Nov 29 10:44] Setting affinity to run helper thread 3 on logical CPU #3
[Work thread Nov 29 10:44] Using FFT length 20M, 4 threads
[Work thread Nov 29 10:44] p: 332190991.  Time: 241.513 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 323.384 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 226.639 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 300.505 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 225.736 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 228.671 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 234.370 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 225.403 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 224.832 ms.
[Work thread Nov 29 10:44] p: 332190991.  Time: 243.857 ms.
[Work thread Nov 29 10:44] Iterations: 10.  Total time: 2.475 sec.
[Work thread Nov 29 10:44] Estimated time to complete this exponent: 951 days, 13 hours, 17 minutes.
[Work thread Nov 29 10:44] Worker stopped.

this is with a Q6600 @ 3.0GHz with 1066MHz RAM
it seems to scale badly past 2 threads

[gd_barnes]

Originally Posted by jinydu

Well the question did say "current". As I mentioned in a later post, I'm only asking about tests registered on or before 2008. So there isn't much time left to future speed increases.

Still, you're right that the estimated time to complete a 100M digit test is still well within a human lifetime. I think it's mainly a question about human patience, although another consideration is the chance that the computer could fail before the test finishes. I don't know whether any of the users who have registered such a test understand how long it will take, and are sincerely intending to carry through with it anyway. For instance...



As for your question about a billion digit test... So far that isn't even possible on Prime95 because it only supports test up to M999,999,999.

Ah, OK. Thanks for the additional explanation. I guess I didn't read the original statement quite correctly. I see now that you specified "current" tests. Yes, it's all about patience and a reliable machine.

One thing about a reliable machine though: I don't think it would create a big problem in most situations if the machine failed. It depends on how it failed. That's because LLR saves off where it is at at whatever interval you tell it to by creating that little file that starts with a 'z' (usually). So even if the processor or motherboard burned up, as long as the hard drive was still OK, you could swap the hard drive into another good machine or otherwise copy the 'z' file to another machine and continue the work.

I think the only hardware failure to stop a test cold in it's tracks is a hard drive crashing where the user had not backed up the 'z' file (or whatever the file would be for such tests).

Moral of the story: If you are embarking on such an endeavor, make sure to back up the file that it creates once/day or every few days so that you don't potentially lose weeks/months/years of work!


Gary

[gd_barnes]

Originally Posted by James Heinrich

Will one of the current ~480 assigned tests be finished by the user who currently has it reserved? I say very probably not. Will a 100M test be finished in the next 5 years or so? Almost definitely. 5 years from now the per-core processing power will likely be ~10x the Core2 level; moreover 32-core CPUs will almost certainly be around by then. Assuming those two predictions are true then the current 5+ year test will turn into a 1-week test on then-current hardware.

But you can't do one test on multiple cores can you? If so, I want to know about it! (lol) Sure, you could do multiple tests and complete them all at once giving your calculated increase in overall computing capacity, but your increase in the speed of one specific test would still be limited to the increase in speed of the CPU, not the increase in computing capacity.


Gary

[James Heinrich]

Originally Posted by gd_barnes

But you can't do one test on multiple cores can you? If so, I want to know about it!

Sure you can! Any of the v25.x Prime95 clients are multithreaded, you can have it work on either several different assignments, or run one assignment multithreaded, as you see fit. (Currently TF work is not multithreaded, that may change in the future.)

[henryzz]

Originally Posted by gd_barnes

But you can't do one test on multiple cores can you? If so, I want to know about it! (lol) Sure, you could do multiple tests and complete them all at once giving your calculated increase in overall computing capacity, but your increase in the speed of one specific test would still be limited to the increase in speed of the CPU, not the increase in computing capacity.


Gary

you have used llr for too long
one of these day i would like to have a llr program that run one test per core from the same input file