Productivity
Message boards :
Chat :
Productivity
Message board moderation
Author | Message |
---|---|
Send message Joined: 26 Jun 08 Posts: 640 Credit: 437,822,398 RAC: 108,999 |
No comments only because I haven't had a chance to really think about what would be required for a kilobit number. I just downloaded the M1039 paper. For sieving, they used 16e with FB limits of 300M on both sides, which used about 1GB of memory. This shouldn't be a problem. They used 38-bit LPs, which seems a bit excessive. I would test with 33-bit or 34-bit LPs first. This would also help significantly with the filtering. It would reduce the number of relations for a "good" matrix to 1-2 billion. I don't suspect filtering would be too much of a problem. I have a computer available with plenty of RAM for that. Linear algebra is what I fear. They ended up with a 66.7M matrix. This would take more than a year using the BL in the current version of msieve. I have tried using the CADO BW, but I have not gotten good scaling with it either beyond 16 cores. I'm not sure there are publicly available tools that can solve this matrix in less than a year on the computers I have available. I can write a Teragrid research proposal for the computer time if necessary, but that doesn't help without a good implementation. And, of course, the square roots shouldn't be a problem. Ideas for solving the LA issue? |
Send message Joined: 5 Sep 09 Posts: 1 Credit: 49,892 RAC: 0 |
Linear algebra is what I fear.Does the latest msieve solve any of these problems? |
Send message Joined: 2 Oct 09 Posts: 50 Credit: 111,128,218 RAC: 0 |
Linear algebra is what I fear.Does the latest msieve solve any of these problems? I'm not sure that we know the limits of parallel Block Lanczos; but all of the recent records have been set with LA done by block Wiedemann (RSA200, SNFS1024, RSA768). Already the target 5, 409- C282 may be in "big iron" range, with a request for substantial time on first-rate hardware (infiniband ...). The most recent records have relied on distributing the matrix, with each piece run in parallel, which is a Wiedermann feature. -Bruce |
Send message Joined: 26 Jun 08 Posts: 640 Credit: 437,822,398 RAC: 108,999 |
Does the latest msieve solve any of these problems? Up to at least a kilobit SNFS, I believe so. But I want to find out for sure, hence the increased importance of and credit boost for 16e sieving. Recruit all your friends whose computers have sufficient memory, and even your enemies! |