Newer milestone thread - Page 85

flashjh · 2013-12-02, 23:46

Sorry, eating dinner. Yes, but remember that the GPU sieve doesn't work for less than 2^64 right now. Oliver plans to integrate all the 'repetitive' code from the kernels into one function (or something like that) so that mfaktc supports GPU sieve at all bit levels. I've also asked for a MORE_MORE_CLASSES version, if you will, for when we get to very large iteration times so that we don't have to wait so long for a save file and a screen update. All possibilities in the future.

Edit: What is the 'typographic glitch on the screen'?

LaurV · 2013-12-03, 02:16

Quote:

Originally Posted by petrw1

My 2 cents ... that small range could be cleared faster by a big GPU than the time it takes to talk about it. I vote we keep GPU72 more focused and less special cases.

I vote with this. I don't want any "special range" for now, and I only work these small expos when I need my CPU/GPU to do other (real life work) things. Mfaktc does not work optimally on such small expos/bitlevels, because it can't use the "_gs" kernels (gpu sieve), it will use kernels without "_gs" which sieve on CPU. The disadvantage is that you only get half of the "credit" the card can do, and you also waste 1-2 CPU cores, but the advantage is that you only use half of the resources, you produce less heat, consume less electricity

and your system can rotate PCB's and acad parts with reasonable speed...

Edit: re: using v0.18 or v0.19: that affirmation I did for exponents under 1M, which we were testing that time. Here we are talking about exponents of 2M+, so v0.20 can be used but for such exponents, it will only do GPU sieving only from 65 bits up (see my previous posts, and the first paragraph of this current post). In fact, the "less classes" version (not to be confused with "lowest expo" version) of mfaktc would be theoretically the best choice: it has to be the fastest (about 10-20% faster!) for so "short time" assignments, due to lower overhead with the management of the classes.

ET_ · 2013-12-03, 07:58

Quote:

Originally Posted by flashjh

Edit: What is the 'typographic glitch on the screen'?

The program says "class xxx/4620" even if it computes only 420 classes.

Luigi

ET_ · 2013-12-03, 08:02

The first 100 expos of the list are done, the results attched.

Who's going to go on?

Luigi

LaurV · 2013-12-03, 09:59

Quote:

Originally Posted by ET_

The program says "class xxx/4620" even if it computes only 420 classes.

Luigi

And you can't fix this from the ini file? (there were plenty of options for printing there, you can also use numerical in print format).
Don't ask me how I know for a long time!

ET_ · 2013-12-03, 10:22

Quote:

Originally Posted by LaurV

And you can't fix this from the ini file? (there were plenty of options for printing there, you can also use numerical in print format).
Don't ask me how I know for a long time!

Gotcha!

BTW, did playing with NumStreams and CpuStreams add efficiency? I could try myself, but I'm so lazy...

L.

LaurV · 2013-12-03, 13:19

First victory!

Code:

FACTORS FOUND: 1 
PROCESSING RESULT: M2186747 HAS A FACTOR: 6816375925079296039 
[TF:62:63*:MFAKTC 0.20 75BIT_MUL32] 
CPU CREDIT IS 0.5339 GHZ-DAYS.

ET_ · 2013-12-03, 13:39

Quote:

Originally Posted by LaurV

First victory!

Code:

FACTORS FOUND: 1 
PROCESSING RESULT: M2186747 HAS A FACTOR: 6816375925079296039 
[TF:62:63*:MFAKTC 0.20 75BIT_MUL32] 
CPU CREDIT IS 0.5339 GHZ-DAYS.

Congrats!

I'll take 100 more for tonight.

Luigi

bloodIce · 2013-12-03, 17:09

Quote:

Originally Posted by LaurV

First victory!

Code:

FACTORS FOUND: 1 
PROCESSING RESULT: M2186747 HAS A FACTOR: 6816375925079296039 
[TF:62:63*:MFAKTC 0.20 75BIT_MUL32] 
CPU CREDIT IS 0.5339 GHZ-DAYS.

Congratulations, LaurV! One down

. If you finish these 700 and ET crunches the 200, then more than a quarter of the work is done. It would have taken me above two months continuous dedication to TF 900 of those expos. And to both of you it took less than 24 hours as I see. Good work guys

.

lycorn · 2013-12-03, 17:10

I would like to propose that we collectively engage in an effort to take to 65 bits all exponents that are below that level.
I have identified 3 different situations:

1. Exponents lower than 1 MB, regardless of the current bit level they are factored to:
These cases have to be dealt with using the mfaktc version that works with small exponents. All caveats mentioned by TheJudger and LaurV in this thread shall be taken into account.

2. Exponents > 1MB, factored to 63 bits or less:
GPU sieving doesn´t work for these bit levels, so either it is disabled in the config file or we use 0.19 or lower.

3. Exponents > 1 MB, and factored to 64 bits:
For these exponents, mfaktc 0.20 shall be used.

For all the above situations, CPUs may obviously be used. This might interest people with old hardware/32-bit OS. Old Athlons, or Intel Core architecture, are good examples.

GPUto72 could control the flow of assignments/reports, as suggested above in the thread.
While that is not implemented, if it will ever be, the exponents may be obtained by copying them from appropriate "Factoring Limits" Primenet pages, and the manual test pages may be used to report results. Also, during this interim (pre GPUto72) phase I would urge all participants in this quest to reserve all assignments they are working in, and to only request non-assigned exponents, so we don´t step on each other toes.

Opinions? Suggestions? Critics (just spare me of "It´s of no use for GIMPS purposes")?

P.S.: I´m currently doing some 2.52 M to 63. Please be so kind as to...

P.S. 2: Suggestions on what to do about cracking the three tough nuts below 2K also welcome. There are currently no ways to put a GPU at it, and even CPU TF will not work (at least from Prime95 version 25 onwards). Let´s just sit and wait for SNFS to do it, or?...

flashjh · 2013-12-03, 17:11

I'm starting on 2347031. That's where you left off, right? (after the 100 for tonight)

2013-12-02, 23:46	#925
flashjh "Jerry" Nov 2011 Vancouver, WA 1,123 Posts	Sorry, eating dinner. Yes, but remember that the GPU sieve doesn't work for less than 2^64 right now. Oliver plans to integrate all the 'repetitive' code from the kernels into one function (or something like that) so that mfaktc supports GPU sieve at all bit levels. I've also asked for a MORE_MORE_CLASSES version, if you will, for when we get to very large iteration times so that we don't have to wait so long for a save file and a screen update. All possibilities in the future. Edit: What is the 'typographic glitch on the screen'? Last fiddled with by flashjh on 2013-12-02 at 23:47

2013-12-03, 13:19	#931
LaurV Romulan Interpreter Jun 2011 Thailand 2⁶·151 Posts	First victory! Code: FACTORS FOUND: 1 PROCESSING RESULT: M2186747 HAS A FACTOR: 6816375925079296039 [TF:62:63:MFAKTC 0.20 75BIT_MUL32] CPU CREDIT IS 0.5339 GHZ-DAYS. Last fiddled with by LaurV on 2013-12-03 at 13:20*

2013-12-03, 17:11	#935
flashjh "Jerry" Nov 2011 Vancouver, WA 1,123 Posts	I'm starting on 2347031. That's where you left off, right? (after the 100 for tonight) Last fiddled with by flashjh on 2013-12-03 at 17:19

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Newer X64 build needed	Googulator	Msieve	73	2020-08-30 07:47
Performance of cuda-ecm on newer hardware?	fivemack	GMP-ECM	14	2015-02-12 20:10
Cause this don't belong in the milestone thread	bcp19	Data	30	2012-09-08 15:09
Newer msieves are slow on Core i7	mklasson	Msieve	9	2009-02-18 12:58
Use of large memory pages possible with newer linux kernels	Dresdenboy	Software	3	2003-12-08 14:47

2013-12-03, 17:10	#934
lycorn "GIMFS" Sep 2002 Oeiras, Portugal 2·11·67 Posts	I would like to propose that we collectively engage in an effort to take to 65 bits all exponents that are below that level. I have identified 3 different situations: 1. Exponents lower than 1 MB, regardless of the current bit level they are factored to: These cases have to be dealt with using the mfaktc version that works with small exponents. All caveats mentioned by TheJudger and LaurV in this thread shall be taken into account. 2. Exponents > 1MB, factored to 63 bits or less: GPU sieving doesn´t work for these bit levels, so either it is disabled in the config file or we use 0.19 or lower. 3. Exponents > 1 MB, and factored to 64 bits: For these exponents, mfaktc 0.20 shall be used. For all the above situations, CPUs may obviously be used. This might interest people with old hardware/32-bit OS. Old Athlons, or Intel Core architecture, are good examples. GPUto72 could control the flow of assignments/reports, as suggested above in the thread. While that is not implemented, if it will ever be, the exponents may be obtained by copying them from appropriate "Factoring Limits" Primenet pages, and the manual test pages may be used to report results. Also, during this interim (pre GPUto72) phase I would urge all participants in this quest to reserve all assignments they are working in, and to only request non-assigned exponents, so we don´t step on each other toes. Opinions? Suggestions? Critics (just spare me of "It´s of no use for GIMPS purposes")? P.S.: I´m currently doing some 2.52 M to 63. Please be so kind as to... P.S. 2: Suggestions on what to do about cracking the three tough nuts below 2K also welcome. There are currently no ways to put a GPU at it, and even CPU TF will not work (at least from Prime95 version 25 onwards). Let´s just sit and wait for SNFS to do it, or?...