mersenneforum.org

mersenneforum.org (https://www.mersenneforum.org/index.php)
-   NFS@Home (https://www.mersenneforum.org/forumdisplay.php?f=98)
-   -   2018 15e post-processing reservations and results (https://www.mersenneforum.org/showthread.php?t=23411)

swellman 2018-12-13 22:05

[QUOTE=Batalov;502647]The initial seeding for L2875B and L2815A appears to be short-ish.
I am more concerned about L2875B which is spec'd as a straight 32/32 job.

L2815A might work because it is a 32/30 job. I will test it when it will reach 99% received for the initial push.[/QUOTE]

Just bumped L2875B upper Q limit to 300M. The recommendation by NFS is 336M, so I can press it higher later if you wish.

I can bump L2815A higher if you find it undersieved, or I can do it now if you would prefer. NFS recommends raising the upper limit of Q to 375M! Of course the recommendation is not always correct, especially on 32/30 3LP jobs...

fivemack 2018-12-14 17:24

Remember that L2875B and L2815A are quartics, which in my experience sieve relatively slowly but do not require all that many relations to give a convenient matrix. I would have been strongly inclined to run at least a first filtering quite early, with 400M-ish relations, but I have enough hardware that even a 30M matrix takes a week or so, and I am basically indifferent between a matrix job that takes two days and a matrix job that takes a week.

(wget -c works well for these jobs, so it's just a matter of an incremental download if the first filtering doesn't produce something usable)

L2043 I expect to be a bit more of a challenge; I'm downloading now but it may well want rather more relations since it's an SNFS 285.

Batalov 2018-12-15 00:30

One of those days...

[STRIKE]One[/STRIKE] Two of 'my' dat.gz files have broken content in the middle (well, rather near beginning), so I will have to [URL="https://github.com/arenn/gzrt"]repair it[/URL] by dumping in chunks and then collecting back again.

Is there a validation test while receiving the end-of-task content? Maybe a lightweight test could be added (not a full zcat but maybe just a gz-header checker)? There are might be spoofers on the loose in the NFS @ Home pool.

fivemack 2018-12-15 08:22

[QUOTE=Batalov;502822]One of those days...

[STRIKE]One[/STRIKE] Two of 'my' dat.gz files have broken content in the middle (well, rather near beginning), so I will have to [URL="https://github.com/arenn/gzrt"]repair it[/URL] by dumping in chunks and then collecting back again.

Is there a validation test while receiving the end-of-task content? Maybe a lightweight test could be added (not a full zcat but maybe just a gz-header checker)? There are might be spoofers on the loose in the NFS @ Home pool.[/QUOTE]

My suspicion is that the broken content is due to a disc-full condition on the server rather than to malice - I’ve had to gzrecover several files from around the time that the Lucas sieving paused for weeks.

Batalov 2018-12-15 16:51

Quite possible. The third file is ok.

fivemack 2018-12-15 21:03

As of when I downloaded it, L2043 has 319309673 unique relations and produces a 53.1M density-70 matrix with an ETA of 870 hours (256-bit vectors, 10 cores Xeon Silver 4114). I will leave msieve crunching overnight to see what higher densities bring (102 doesn't work, 49.07M density 86), but have asked the grid to sieve 600M-800M.

fivemack 2018-12-17 13:32

C250_142_122 completed
 
1 Attachment(s)
[code]
Sun Dec 16 20:14:24 2018 p64 factor: 3379282931282729067965022974297849754323663102589254017059459601
Sun Dec 16 20:14:24 2018 p77 factor: 10004118294661585092031906998707324054788436039714785889620138503417492957169
Sun Dec 16 20:14:24 2018 p111 factor: 194454949893438217095267096107091674631220490806841113847109503137317801838748423991807657786341228278304330921
[/code]

209.6 hours for 21.76M density-132 (not 134) matrix on 6 cores i7/4930K, la_block=8192, VBITS=128.

Log attached and at [url]https://pastebin.com/gz2tEarN[/url]

fivemack 2018-12-20 12:30

5039_73m1 completed
 
1 Attachment(s)
[code]
Thu Dec 20 06:04:46 2018 p71 factor: 20053882495885784345674750052933475924857668614431355963762764026282631
Thu Dec 20 06:04:46 2018 p80 factor: 91103441492973369500334775256951724098155980467477580956345591982274679892569183
Thu Dec 20 06:04:46 2018 p117 factor: 202836798815734208960835393388441000650910050237709972994172229972380207457398369544572746718675248714254223293572457
[/code]

147.6 hours on 14 cores i9/7940X VBITS=256 for 25.38M matrix of density 124 (not 126). Log attached and at [url]https://pastebin.com/fYhRWVXe[/url]

Batalov 2018-12-21 03:51

L2815A is done
[CODE]Thu Dec 20 19:34:08 2018 p81 factor: 243210235876352865132563116819787566290132639076456130191070988073047492192020261
Thu Dec 20 19:34:08 2018 p154 factor: ...
Thu Dec 20 19:34:08 2018 elapsed time 00:39:18
[/CODE]The 32/30 parametrization is the way to go. L2875 has a harder fate with the straight 32/32


[PASTEBIN]MFSjB8ht[/PASTEBIN]

fivemack 2018-12-21 18:19

[QUOTE=fivemack;502922]As of when I downloaded it, L2043 has 319309673 unique relations and produces a 53.1M density-70 matrix with an ETA of 870 hours (256-bit vectors, 10 cores Xeon Silver 4114). I will leave msieve crunching overnight to see what higher densities bring (102 doesn't work, 49.07M density 86), but have asked the grid to sieve 600M-800M.[/QUOTE]

With the extra sieving, I have 390795360 unique relations, and the density-70 matrix is now 39.64M; not enough relations for density 134, but this seems entirely tractable.

Eventual matrix 32.15M at density 130, ETA 257 hours on 14 cores i9-7940X. Yes, it's a faster computer, but eleven days instead of 36 makes the extra sieving seem worth it in a way that usually I haven't seen.

I'm on vacation over New Year so expect factors January 4 or 5.

fivemack 2018-12-21 18:44

C252_127_119 complete
 
1 Attachment(s)
[code]
Fri Dec 21 13:18:27 2018 p118 factor: 1214128006544997396906395373792921681209637127710644614411333640468297717854995607781720784226638937031747887594648223
Fri Dec 21 13:18:27 2018 p135 factor: 165051426409474495671739620717188684143775531081428424189795285122214109748648854239804327059802568645710005284291274641506950067461379
[/code]

259.8 hours on 4 cores i7-4790K (la_block=4096, VBITS=128) for 16.76M density-144 matrix; comparative advantage would have me leave the 4-core machines (possibly the relevant fact is dual-channel-memory machines) to aliquot sequence work and run all the linear algebra on Sky Lake.

Log attached and at [url]https://pastebin.com/ayyEy0SP[/url]


All times are UTC. The time now is 03:06.

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2021, Jelsoft Enterprises Ltd.