A computation worthy of the name

[fivemack]

I have spent a little while figuring out parameter choice for this; here is the polynomial for 2^941 - 1:

n: 2468815427808777830385945175570959888649322954010151512577652841207233127175594988619285951813249353482864131674513502872982111245757912252275917421142241044267118542381889815949472721865947896251559297022100274282885756253947868389157090816092206938192397903861973556952725044519
c6: 1
c0: -2
Y0: 182687704666362864775460604089535377456991567872
Y1: 1
skew: 1.12
rlim: 300000000
alim: 250000000
lpbr: 32
lpba: 32
mfbr: 64
mfba: 64
rlambda: 2.6
alambda: 2.6

We will need about six hundred million raw relations, which I propose to get by sieving with -16e from Q=50 million to somewhere a little past Q=200 million on both sides. I have done 0.005% of the job (1k in each 20M) on a quad-core 2.4GHz Core 2 running 64-bit Linux, with 8G of memory and 16G of swap; it took from 8:30pm to half past midnight, and used no swap space.

Each CPU-minute produces about forty relations; fifteen million minutes is thirty years. This job would take a day even on one of the pride-of-the-National-Labs supercomputers; I'll need all the help I can cajole.

The final parts of the job will take me about a season on a reasonably fast Core i7 system.

RESERVATIONS CLOSED 24/11

RESERVATIONS
10/9 andi47 50-51 A+R (A side done 30/10, R side partially handed over to fivemack and done 13/11)
10/9 bsquared 51-61 A+R done 29/9
10/9 fivemack 61-67 A+R done 25/10 (slowed by uploading issues)
10/9 batalov 67-69 A+R done 29/9
01/10 bsquared 69-74 A+R done 11/10
13/10 J.F. 74-75 A+R done 22/11
13/10 bsquared 75-80 A+R done 26/10
15/10 fivemack 80-86 A+R done 13/11
22/10 frmky pp NFS@home 86-90 A+R done 11/11
24/10 frmky pp NFS@home 90-100 A+R done 11/11
30/10 bsquared 100-102 A+R done 06/11
08/11 fivemack 102-107 A+R done 24/11
13/11 bsquared 107-110 A+R 97.5%-done 19/11, finished 23/11
10/9 bdodson 110-190 A+R done 20/11
17/11 bdodson 190-200 A+R
16/11 NFS@home 200-225R

RESULTS
18/09 batalov 67-68R 2,122,258r in 2.0Msec PhenomII/3000
22/09 batalov 67-68A 2,156,943r in 2.0Msec PhenomII/3000
25/09 fivemack 61-62A+R 4,340,480r in 5.3Msec various systems
26/09 batalov 68-69A 2,137,457r in 2.0Msec PhenomII/3000
29/09 batalov 68-69R 2,118,321r (presumably in 2.0Msec PhenomII/3000)
29/09 bsquared 51-61A+R; 43,716,577r in 61.1Msec k8/2000.
06/10 fivemack 64-65A+R; 4,300,156r in 5.6Mtsec on eight threads of i7@2800MHz
11/10 bsquared 69-74A+R; 21,206,954r in 31.2Msec k8/2000
(ran filtering job)
13/10 fivemack 63-64A+R; 4,307,838r in 5.0Msec various systems
13/10 fivemack 62-63A+R; 4,313,909r in 4.9Msec C2/2400
15/10 fivemack 65-66A+R; 4,309,003r in 5.6Mtsec on eight threads of i7@2800MHz
25/10 fivemack 81-82A+R; 4,172,590r in 2.45Msec K10/2500 + 2.64Msec C2/2400
25/10 fivemack 66-67A+R; 4,296,209r in 5.05Msec various systems
26/10 bsquared 75-80A+R; 20,068,462 relations in 30.6Msec k8/2000
27/10 bdodson 110-150R; 78,785,145 relations
28/10 fivemack 80-81A; 2077128 relations in 2.7Msec various systems
28/10 bdodson 150-170R; 37,937,272 relations
(ran filtering job)
30/10 andi47 50-51A; 2,265,218 relations
02/11 fivemack 80-81R; 2,095,566 relations in 3.1Msec various systems
02/11 fivemack 82-83A+R; 4,162,666 relations in (a)2.56Msec K10/2500 + (r)2.45Msec C2/2400
04/11 fivemack 83-84A+R; 4,149,546 relations in (a)2.82Mtsec i7/2800 + (r)2.76Mtsec i7/2800 - 241 hours realtime
06/11 bsquared 100-102A+R; 8,076,444 relations
06/11 bdodson 170-190R; 35,163,889 relations
06/11 bdodson 110-130A; 38,603,445 relations
10/11 fivemack 84-85A: 2,062,050 relations in 2.84Msec various systems
10/11 fivemack 84-85R: 2,074,375 relations in 2.43Msec K10/2500
11/11 nfs@home 86-100R: 28,995,972 relations
11/11 nfs@home 86-100A: 28,868,800 relations
13/11 fivemack 85-86A+R; 4,155,760 relations in (a)2.86Mtsec i7/2800 + (r)2.77Mtsec i7/2800 - 204 hours realtime
(ran filtering job: 35.3 megacycles)
13/11 andi47 50-50.2R + fivemack 50.2-51R; 2,153,725 relations
13/11 bdodson 130-170A: 72,248,902 relations
(ran filtering job; 28.1M matrix)
17/11 fivemack 102-103A+R: (a) 2,001,779rel in 2.55Msec C2/2400 (r) 2,038,855rel in 2.47Msec K10/2500
18/11 fivemack 104-105A+R: (a) 1,981,674rel (r) 2,030,985rel (various systems)
19/11 bsquared 107-110A+R (97.5% done) 11,709,149 rel; last relations collected 23/11
20/11 bdodson 170-190A: 35,387,263 rel
22/11 fivemack 103-104A+R: (a) 1,988,386rel in 2.82Mtsec i7/2800 + (r) 2,034,560 rel in 2.78Msec i7/2800 - 204 hours realtime
22/11 J.F. 74-75A+R: 4,197,092 rel, 5.02Msec runtime
23/11 bdodson 190-200R + patches from 160A and 180R: 18,329,135 + 1,328,828 + 1,883,125 rels
24/11 fivemack 105-106A+R: (a) 1,988,962 rel in 2.60Msec K10/2500 + (r) 2,026,728 rel in 2.46Msec C2/2400
24/11 fivemack 106-107A+R: (a) 1,950,983 rel in 2.69Msec various + (r) 2,024,601 rel in 2.45Msec various

[fivemack]

measured over a 1k range and extrapolated idealwise

[fivemack]

wouldn't it be nice to attach multiple files to the same post?

[fivemack]

Looks like about three CPU-days for a 100k range, so this is a factor three or so slower than 2^877-1 was

[Andi47]

I assume that this one will run for several months, so there might be enough time to run it over various weekends (even on a system with "only" 2 GB), so I will reserve a chunk which needs least memory (due to necessarily low factorbase bound(s)):

reserving 50M - 51M (r+a)

I will run it on only one core of a Core 2 Duo (Win XP 32 bit, 1 GB RAM; testing needed ~800MB Ram and 1.6 GB virtual memory) and will run only when I am out of office (weekends and so, with the 2nd core doing a low-memory-job like GIMPS TF'ing), so it might take quite a while.

If it turns out that it does too much swapping (i.e. slowly murdering the harddisk), I might stop and unreserve somewhere in the middle.

Edit: I will start in ~1 week to 10 days, when my current c153 (2,1766M) finishes sieving (this one was at ~45.1M rels, 39.7M unique, when I did a filtering run on thuesday).

[Andi47]

Suggestion: How about rlim=268435455 (2^28-1)? This might be somewhat more memory efficient.

fivemack: Yes, quite possibly; try it if you want and report if it saves anything perceptible.

[Batalov]

Damned if I ever passed a good 1% stake in a computation worthy of the name! Count me in for the treacherous ranges 67-69M a+r.

_____
I'll stand under the bridge I laid my hands on.

[bsquared]

Looks like the university's 36 k8/1400 cpus are not available now, but happily 40 k8/2000 cpus are... so I'll take 51M-61M A+R.

- ben.

[Andi47]

Originally Posted by Andi47

Suggestion: How about rlim=268435455 (2^28-1)? This might be somewhat more memory efficient.

fivemack: Yes, quite possibly; try it if you want and report if it saves anything perceptible.

OK, I will give it some test runs (today or tomorrow) (just for some 2000 Q's to measure the memory useage).
As I said, I will start the "productive" run in approx. 7-10 days, namely when 2,1766M finished sieving.

[bdodson]

Originally Posted by bsquared

Looks like the university's 36 k8/1400 cpus are not available now, but happily 40 k8/2000 cpus are...
- ben.

Slow here too. I'll take 130M-190M-(a+r). I'll start after the last
two of our marvelous seven firsts finish (1642L should factor today,
the c169 gnfs on Friday, 5p346 should finish sieving sometime today,
11p233 will start today, last one is 5m347). -Bruce

[Andi47]

sieving with 16e on the a side from Q=50M to 50M+200 on a P4 (win XP, 32 bit)

rlim = 300M
max. mem = 874 MB
virtual mem. = 1.749 MB
total yield: 68, q=50000201 (3.13947 sec/rel)

rlim=2^28-1:
max. mem = 825.7 MB
virtual mem = 1.666 MB
total yield: 66, q=50000201 (3.13068 sec/rel)

(can please someone verify the timing, hence mine was done on a somewhat busy P4?)