A computation worthy of the name

[bsquared]

Originally Posted by bsquared

Reserving 69-74M, A+R

Done and uploaded.

A side: 10644822 relations, 15.9Msec (k8/2000)
R side: 10562132 relations, 15.3Msec (k8/2000)

[fivemack]

Added to the collection; thanks!

(I am making forward progress, though it's on a distributed pile of systems so not very uniform forward progress; should finish my reservation by middle of next week)

[fivemack]

Time since project started: 32 days

Relations in /home/nfsworld/2-941/relations: 19 1MQ A+R ranges (so perhaps 13% of the way to the target, ETA middle of May 2010 but I'm not accounting for Bruce's resources coming into play), 5353MB of .gzip files containing 82.1M relations.

Mon Oct 12 20:04:46 2009  commencing duplicate removal, pass 1
Mon Oct 12 20:28:44 2009  found 8416013 hash collisions in 82091502 relations
Mon Oct 12 20:29:14 2009  added 1219003 free relations
Mon Oct 12 20:32:02 2009  found 6575925 duplicates and 76734580 unique relations
Mon Oct 12 20:59:05 2009  begin with 76734580 relations and 152720099 unique ideals
Mon Oct 12 21:03:01 2009  reduce to 262 relations and 0 ideals in 3 passes

which is quite a high duplicate rate, but I think better than the one Bruce and I saw using 15e sieving for 12^256+1.

About one relation in 100,000 has a b value greater than 2^32, which provides a nice running progress report as the msieve run starts.

[J.F.]

Thanks to Fivemack, I now have a fresh working 16e executable :). Reserving 74-75M A+R, just to get things up and running.

[bdodson]

Originally Posted by fivemack

Time since project started: 32 days

Relations in /home/nfsworld/2-941/relations: 19 1MQ A+R ranges (so perhaps 13% of the way to the target, ETA middle of May 2010 but I'm not accounting for Bruce's resources coming into play), ...

We're doing 50M-200M-(A+R), for 150M-times-two, and I have 80M-times-two
reserved; for a bit past half of the first estimate of the range needed. The
other half is 70M, so does 19M/70M = 27% sound like it's feasible for the
forum to complete the other half if I get stuck? Suppose it's plausible.

In any case, doesn't sound like Greg is in any hurry for the R269 relations
to show up, and I'm planning a break to run the -R half of my half. Just so
there's not any doubt that I'm still in. -Bruce

(PS - I'm just finishing 100M-190M-R with 16e on R269, so I have my
scripts for 110M-190M-R wired. That's half of the 20M-200M-R range
that Greg set for R269 = 111....11 (269 1's).)

[bsquared]

Reserving 75-80 A+R. This range may take longer as some of the cluster nodes are busy and I'll have to wait for the queue to clear.

[fivemack]

Originally Posted by bdodson

We're doing 50M-200M-(A+R), for 150M-times-two, and I have 80M-times-two
reserved; for a bit past half of the first estimate of the range needed. The
other half is 70M, so does 19M/70M = 27% sound like it's feasible for the
forum to complete the other half if I get stuck? Suppose it's plausible.

It's entirely plausible, the sieving for other projects has extended over more than a hundred days without everyone getting bored and dropping out. J.F. and bsquared both have access to reasonably significant resources; I could complete the other half in a year unassisted, though I'd rather not.

[bdodson]

Originally Posted by fivemack

It's entirely plausible, the sieving for other projects has extended over more than a hundred days without everyone getting bored and dropping out. J.F. and bsquared both have access to reasonably significant resources; I could complete the other half in a year unassisted, though I'd rather not.

110M-120M-R is running. On R269 with 16e, 110M-190M-R took under
two weeks, even with moderate loading on the clusters. Suppose
M941 will run more slowly(?). -bd

[fivemack]

I'd expect M941 to be a bit slower than R269; if you PM me the polynomial for R269 I'll run a comparison on a machine whose M941 performance I know, but otherwise I think the data point is that a 10M A+R range, with significantly smaller Q, took 60 million CPU-seconds for bsquared's cluster.

Thanks again for the cycles!

[bdodson]

Originally Posted by fivemack

I'd expect M941 to be a bit slower than R269; if you PM me the polynomial for R269 I'll run a comparison on a machine whose M941 performance I know, ...

I found two ranges of width 1M (run in 250 tasks of width 4K) for which
neither R269 nor M941 ran on the small memory Opterons --- 110M for one
119M for the other. Condor reports the number of hours to complete the
4K tasks as

cat fgo-alt269m-119*log | grep 'Run Remote' | cut -c 9-11 | sort | uniq -c
     28 02:
     37 03:
    185 04: 

cat fgo-alt941m-111*log | grep 'Run Remote' | cut -c 9-11 | sort | uniq -c
     41 02:
     25 03:
    178 04:
      6 05:

At the risk of data overload, by number of 10's of minutes

cat fgo-alt269m-119*log | grep 'Run Rem' | cut -c 9-12 | sort | uniq -c
      1 02:1
      1 02:2
      6 02:3
      7 02:4
     13 02:5
      9 03:0
      2 03:1
      2 03:2
      5 03:3
      9 03:4
     10 03:5
     29 04:0
     55 04:1
     50 04:2
     35 04:3
     13 04:4
      3 04:5

cat fgo-alt941m-111*log | grep 'Run Rem' | cut -c 9-12 | sort | uniq -c
      4 02:2
     11 02:3
     13 02:4
     13 02:5
     12 03:0
      2 03:1
      2 03:3
      2 03:4
      7 03:5
     18 04:0
     16 04:1
     31 04:2
     44 04:3
     50 04:4
     19 04:5
      6 05:0

which suggests that somewhat more of the M941 tasks ran on fast
xeons (and equally fast Opteron quad cores, Processor 8384, with
each virtual machine taking all four cores). There's an apple-to-apple
comparison among the tasks that ran on the dual-quad-core cluster,
eight virtual machines/ dual quad, 170 of the M941 tasks, 181 of the
R269 tasks. These dual quad timings will be among the times at/above
4.00, with 184, resp. 185 tasks for each. That's

cat fgo-alt269m-119*log | grep 'Run Rem' | cut -c 9-12 | sort | uniq -c
         ...
     29 04:0
     55 04:1
     50 04:2
     35 04:3
     13 04:4
      3 04:5

cat fgo-alt941m-111*log | grep 'Run Rem' | cut -c 9-12 | sort | uniq -c
         ...
     18 04:0
     16 04:1
     31 04:2
     44 04:3
     50 04:4
     19 04:5
      6 05:0

Mmmph. So the R269's are clustered between 4:10-4:29 (105 tasks), while
the M941's are clustered between 4:30-4:49 (94 tasks). An extra 20 minutes
for the M941 tasks, out of 280 minutes? -Bruce

PS -- These timing numbers are consistent with my impression of the comparison
between my two previous 16e numbers, Greg's 5p398 and Serge's 2M2086, the latter
also at difficulty c. 269, like R269. Waiting for 2500 tasks to finish a 10M
range of q's didn't seem notably different for the easier number than for the one
that's more difficult. So most likely not a multiple-precision break for the
siever; of the type that would correspond to substantially harder sieving.
Gradual increase in timing, as distinct from a break.

[frmky]

Do you mind if I take a range for NFS@Home? This will allow me to test the BOINC community response to the 16e sieve without committing to a huge factorization.