[QUOTE=VBCurtis;387075]*snip*
Edit: GPU-ECM has been compiled for 512-bit inputs, 1024-bit, and 4096-bit. I am not sure how large that limit can be compiled for, but at present the max size is hard-coded at compile time.[/QUOTE]

Although it was compiled for these sizes, it doesn't actually work on those numbers. Still trying to figure out how (if I can) to get it to work, but with a new job and lack of actual programming skills, that probably won't happen. In the meantime, the highest working limit is 2^1018-1.

[QUOTE=Madpoo;386869]Well, on that note we're actually trying out some other tweaks. Right now if you show the full details, it will include every time someone checked in even just 1 curve of an ECM run.[/QUOTE]
:direction:
Can all of the ECM and non-milestone discussion get moved to a new thread? Thank you.

[QUOTE=wombatman;387080]Although it was compiled for these sizes, it doesn't actually work on those numbers. Still trying to figure out how (if I can) to get it to work, but with a new job and lack of actual programming skills, that probably won't happen. In the meantime, the highest working limit is 2^1018-1.[/QUOTE]

The reason the limit is 2^1018-1 is because of the size limit for modular operations on the GPU. I don't know if this can be worked around, but it might be able to be.

[SIZE="6"][FONT="Lucida Sans Unicode"]Cinco[/FONT][/SIZE]

[QUOTE=Uncwilly;387100][SIZE="6"][FONT="Lucida Sans Unicode"]Cinco[/FONT][/SIZE][/QUOTE]

...de mayo? You're a bit late, I'm afraid.

[QUOTE=Primeinator;387110]...de mayo? You're a bit late, I'm afraid.[/QUOTE]

That's his way of bringing this thread back on-topic. :)

5 exponents left to prove M44 is really M44.

So, for our next minor milestone, does it seem like the one to double-check all exponents below 10M? Basically baking this in:

[URL="http://www.mersenne.org/assignments/?exp_lo=31148503&exp_hi=33219280&execm=1&extf=1&B1=Get+Assignments"]http://www.mersenne.org/assignments/?exp_lo=31148503&exp_hi=33219280&execm=1&extf=1&B1=Get+Assignments[/URL]

If that seems good, I'll see about adding that to the milestone page maybe today sometime with a link to that report.

[QUOTE=Madpoo;387121]That's his way of bringing this thread back on-topic. :)

5 exponents left to prove M44 is really M44.

[/QUOTE]

My very poor sense of humor. Edit: Or my constant desire for Mexican food.

I also noticed that there are now fever than 6k exponents to test until all numbers below M48 have had one LL.

[QUOTE=Madpoo;387121][URL="http://www.mersenne.org/assignments/?exp_lo=31148503&exp_hi=33219280&execm=1&extf=1&B1=Get+Assignments"]http://www.mersenne.org/assignments/?exp_lo=31148503&exp_hi=33219280&execm=1&extf=1&B1=Get+Assignments[/URL]

If that seems good, I'll see about adding that to the milestone page maybe today sometime with a link to that report.[/QUOTE]That URL only works if all exponents are assigned. There are times when some exponents are unassigned and are missing from the list. A different report would be needed to show all exponents, like the one I mention [url=http://www.mersenneforum.org/showthread.php?p=381855#post381855]here[/url].

[QUOTE=TheMawn;387059]Is the B2 = 100 x B1 convention something that we have since determined to be optimal or is that also strictly for book-keeping?[/QUOTE]

[QUOTE=R.D. Silverman;80379](1) Spending an equal amount of time in step 1 and step 2 maximizes the per-unit time probability of success. This is independent of the step 2 METHOD. If a particular step 2 run K times as fast as step 1, and we take step 1 to B1, then we should take Step 2 to K*B1

In theory, an ordinary FFT implementation [by say a perfect programmer] can take an ordinary step 2 FFT (i.e. without the Brent-Suyama or similar extensions) to B1^2 in the same time it took step 1 to run to B1.[/QUOTE]

B2 = B1^2 would be the optimal with a "perfect" program and "infinite" memory.


[QUOTE=Madpoo;387078]I mean, let's say you have a server with 192GB of RAM and 72 cores... having something like Prime95 doing huge memory tests while also fully thrashing the CPU would really be an awesome burn-in. It'd also exercise the cooling system and dual power supplies. :)[/QUOTE]

I don't think Prime95 can run on 72 cores, but I think 32 cores last I heard. Then you could run GMP-ECM at the same time with a huge B2 to use as much memory as possible:
First you could run stage1 and save at the end: ecm.exe -save stage1.txt 3e6 1 < somenumber.txt
Then you can resume from the savefile to test memory usage by extending B2 and using verbose mode and "-k 1" which ensures all of stage2 is run in 1 step using the most memory:
ecm.exe -v -k 1 -resume stage1.txt 3e6 1e12 (modify 1e12 and check memory usage from the verbose mode output)

I am not sure if GMP-ECM can use as much as 192 Gb, but I guess you could run several copies if not.

[QUOTE=retina;387136]That URL only works if all exponents are assigned. There are times when some exponents are unassigned and are missing from the list. A different report would be needed to show all exponents, like the one I mention [url=http://www.mersenneforum.org/showthread.php?p=381855#post381855]here[/url].[/QUOTE]

In this case, they are all assigned. :smile:

That's one reason I figured this would be the next good milestone to actually have some kind of countdown and a link to a report.

The other existing milestone #'s on that page are all nice, but not all in a range are assigned so without some really wild guessing or some emphasis on assignments, we couldn't do any kind of "expected completion" thing.

Oh well... it's probably nice to have milestones that are more easily digestable, with dates in the not too distant future.

[QUOTE=ATH;387142]I don't think Prime95 can run on 72 cores, but I think 32 cores last I heard.[/QUOTE]
[QUOTE]New features in Version 28.5 of mprime
--------------------------------------

1) Changed the output to the worker windows during LL and PRP tests. The new output includes the estimated time to complete the test. There are two new options described in undoc.txt: ClassicOutput and OutputRoundoff.

2) Added some new options described in undoc.txt: ScaleOutputFrequency, TitleOutputFrequency, and SilentVictoryPRP.

3) Benchmarking on hyperthreaded machines now times only the most common cases. Specifically, hyperthreading is used only in the one cpu and all cpu cases.

4) Benchmarking trial factoring is now off by default. Prime95 should not be used for trial factoring. GPUs are about 100 times more efficient at that task.

5) On multi-core machines, benchmarks are now run on multiple workers. This measures the effect of memory bandwidth during testing and helps you select the setup that gives you the most throughput.

6) There are many new options described in undoc.txt to customize the benchmarking process.

[B]7) Maximum number of threads supported rasied from 64 to 512.[/B][/QUOTE]:mike: