GPU72 and BOINC a match made in ....

[KEP]

Originally Posted by LaurV

You can even split large jobs in classes. I used to do that when factoring my 666M expo to 86 bits, or some 332M to 85. To start from any class is easy, you just create the right checkpoint file (a text editor is enough). To stop at a class, you have to watch (capture the output) and send CTRL+C. You can split the classes in how many cards you have, create 2 or 3 or 4 checkpoint files to resume for the class you want and watch them to stop when the right class is reached. Or, modify mfaktc (not complicate) to stop at some class that he reads from a file similar to the checkpoint file. When you split the job in 4 cards, you do it 4 times faster.

On the other hands, what times are you talking about? It take like 10 minutes to do a 73 and 20 minutes to do a 74 in an average GPU (??), well, more or less, but I do not believe that BOINC community is so sensitive to "10 vs 20 minutes jobs". They may get sensitive when you talk days versus weeks, but when everything is measured in minutes?

You first got me excited about splitting up large jobs in classes over multiple GPUs. This can however not work, using BOINC. The user can not be expected to monitor the progress constantly and then exit BOINC as no more work is needed - let alone exit using CTRL+C. That possibility is not there for the general user to use, if at all.

For test purpose, as long as we try to nail down the errors wunning Linux, a workunit running 5h20 minutes on an average GPU is too long. We simply can not yet accertain how much firepower we have, as long as we have failing workers. Once all the problems is solved, 77-78 is not a problem, simply because then we can almost to the day pinpoint how much work BOINC is able to compute. Patience is our best virtue at the moment

[lalera]

hi,
i think that credit should be given for computing power
for mfaktc the credit you get should be 4 to 5 times higher (estimated)

[LaurV]

Originally Posted by KEP

You first got me excited about splitting up large jobs in classes over multiple GPUs

Do you use "standard" mfaktc? Or you did some special changes to be used in "BOINC world"? If you use the standard one, then I could make you a version (windows only, mfaktc only, no linux, no mfakto, somebody else with more knowledge could adapt it to those) that factors "from class X, to class Y".

[rebirther]

Originally Posted by LaurV

Do you use "standard" mfaktc? Or you did some special changes to be used in "BOINC world"? If you use the standard one, then I could make you a version (windows only, mfaktc only, no linux, no mfakto, somebody else with more knowledge could adapt it to those) that factors "from class X, to class Y".

standard mfaktc, we dont need a split, we are going through the bit lvls, checkpoints are available so all is ok.

[chalsall]

Originally Posted by rebirther

standard mfaktc, we dont need a split, we are going through the bit lvls, checkpoints are available so all is ok.

Read: the easiest to do breadth-first... Exactly the opposite of what I asked you to do. Sigh...

[rebirther]

Originally Posted by chalsall

Read: the easiest to do breadth-first... Exactly the opposite of what I asked you to do. Sigh...

ok, I will add 10k for 77-78 in the next batch so every one is happy then I will go back to 73-74

[KEP]

Originally Posted by rebirther

ok, I will add 10k for 77-78 in the next batch so every one is happy then I will go back to 73-74

Great Reb

[chalsall]

Originally Posted by rebirther

ok, I will add 10k for 77-78 in the next batch so every one is happy then I will go back to 73-74

Thank you!!!

Please understand, this additional firepower is greatly appreciated. However, we are /just/ staying ahead of the FC'ers and P-1'ers. And GPU72 was created to assist GIMPS; we're /not/ the Great Internet Mersenne Factor Search (GIMFS)...

[KEP]

Originally Posted by chalsall

Thank you!!!

Please understand, this additional firepower is greatly appreciated. However, we are /just/ staying ahead of the FC'ers and P-1'ers. And GPU72 was created to assist GIMPS; we're /not/ the Great Internet Mersenne Factor Search (GIMFS)...

We understand ... Now please enlight me a little of the wavefronts and what assignments they are recieving... when Reb returns the 10000 tasks, will it bring us 14 days ahead of the current daily production or how do we infact benefit GIMPs with the 10000 tasks? Is there enough P-1'ed tasks available for first time checkers for all categories according to the assignment rules?

[chalsall]

Originally Posted by KEP

Now please enlight me a little of the wavefronts and what assignments they are recieving... when Reb returns the 10000 tasks, will it bring us 14 days ahead of the current daily production or how do we infact benefit GIMPs with the 10000 tasks? Is there enough P-1'ed tasks available for first time checkers for all categories according to the assignment rules?

OK... As usual, it's complicated...

The first 10,000 candidates Reb has his swarm do will all be without P-1 done. I will then swap out the ~1,000 candidates already at 77 which are currently being given to the P-1'ers, and they should be good for a while. Please note that the P-1 run takes (slightly) less time the higher the candidate has been TF'ed.

The Cat 0 and 1 wavefronts have about 30 days worth of work already appropriately TF'ed and P-1'ed. Cats 2 through 4 are being given work at 77 bits -- Cats 2 and 3 generally with P-1 already done by "specialists", and Cat 4 "just in time" by several of the GPU72 "big guns".

Once the "crunch" is over breadth-first will be fine. But right now a few more runs to 78 (or even just 77) will be appreciated.

Thanks.

[KEP]

Originally Posted by chalsall

But right now a few more runs to 78 (or even just 77) will be appreciated.

Thanks.

I support running everything above 96.83M to 120M to 78 bit. Let's hope that we can use BOINC to push the wavefront of even Category 3 away from "just in time" to actually "sufficient"/"ahead"/"far far ahead"/"never gonna be able to catch us"

Reb, I suggest, that we for the time needed, take up following battleplan:

1. Everything n>96.83M at 77 bit to 78 bit
2. Everything at 76 bit to 77 bit, then everything that remains n>96.83M from 77 bit to 78 bit
3. Everything at 75 bit to 76 bit, then everything that remains at 76 bit to 77 bit and then everything that remains n>96.83M from 77 bit to 78 bit

I know that you and I would both like to go "breadth first" and most likely we can do that from n>110M to n<=150M (maybe n<=200M), once we get far enough ahead of any wavefront. Can you agree on the above suggestion?