Linear algebra reservations, progress and results

[fivemack]

C221_118_81 now running.

Reserving C219_127_57 and C197_129_53 to run over the weekend

[debrouxl]

I'll attempt to queue the OP numbers posted by William probably tomorrow.

While 15e is more efficient in that range (but the 15e queue is full), 14e can usually deal with GNFS difficulty 165-170 tasks, especially with a good poly. SNFS difficulty 250+ with a sextic polynomial on 14e is a stretch.

[xilman]

Originally Posted by debrouxl

I'll attempt to queue the OP numbers posted by William probably tomorrow.

While 15e is more efficient in that range (but the 15e queue is full), 14e can usually deal with GNFS difficulty 165-170 tasks, especially with a good poly. SNFS difficulty 250+ with a sextic polynomial on 14e is a stretch.

Thanks. Indicates where I should devote GPU effort in the next month or two.

Paul

[Dubslow]

Originally Posted by Dubslow

Is there a particular reason for the use of gzip for compression? bzip2 has a rather better compression ratio (at a computational cost) while still being more-or-less widely available.

Originally Posted by frmky

The results are compressed by the BOINC clients before they are returned to the server. The server just checks that they are valid compressed relations then concatenates them into the file you download.

Furthermore, removing duplicates would also be a massive savings of bandwidth. It would perhaps require uncompressing, unless someone extended remdups4 with zlib, but in that case it would become practical to use bzip2.

I'm bringing this up because I have substantially worse internet that I've had years past; it took on the order of 12 hrs to download 22GiB of just over 400M relations, of which 63M were duplicates (and so a waste of bandwidth). Besides the connection being slower, it's also a connection where the total bandwidth consumption is monitored -- and 22 GiB is not insubstantial. (It definitely didn't help that I messed up and needed to do it *again* -- but that was my fault )

[fivemack]

Fri Jan 15 07:37:13 2016  p56 factor: 66604751882840716203190547146724002766120346665371044853
Fri Jan 15 07:37:13 2016  p165 factor: 489911686445026569674165955928329541002751620798887655641477577459794398906029747989127094102551808131327740546125837492604682107848229561628575767032878867858107417

12.8 hours for 5.7M matrix on E5-2650v2 -t 7

[Xyzzy]

Originally Posted by Dubslow

I'm bringing this up because I have substantially worse internet that I've had years past; it took on the order of 12 hrs to download 22GiB of just over 400M relations, of which 63M were duplicates (and so a waste of bandwidth).

It takes us days, but we start downloading relations as soon as they start coming in. At first it is hard to keep up but things balance out nicely towards the end.

while true; do nice wget --continue --limit-rate=64k --user=rsals_data --password=***** http://escatter11.fullerton.edu/nfs_data/12_226_plus_7_226/12_226_plus_7_226.dat.gz; sleep 3600; done

[VictordeHolland]

1373_79_minus1

p63 factor: 167155760887752250734824423685255209540133209961357160033907273
p126 factor: 525234033640980062974735094976085151972433095270924958325303347977625755049053805905136035990038690127663787469942009602039239

12.4M matrix with TD=110
about 109h on all 4 cores 3770k

[VictordeHolland]

Originally Posted by Dubslow

Furthermore, removing duplicates would also be a massive savings of bandwidth. It would perhaps require uncompressing, unless someone extended remdups4 with zlib, but in that case it would become practical to use bzip2.

Bandwidth is not an issue for me (theoretical 90mbit, in practice 50-80mbit depending on the time of day). But need-less to say, if we can limit the download to uniques then I would support that.

However in that case you can't check the duplicate ratio, unless that original number of relations is stored in a table.


[edit]
I''ll be on skiing holidays from tomorrow till Saturday the 23rd. My machines will be off in that time-window.

[pinhodecarlos]

Originally Posted by xilman

These are the sub-S250 remainders:
226.37 7,265- C168
227.22 7,266- C173
227.27 8,249+ C178
227.27 8,249- C173
227.35 6,289- C160
227.44 2,746- C219
227.58 5,322+ C207
227.74 4,374- C215


Paul

Paul,

Let's add them to the queue to help you complete a project milestone . I can thrown two machines to help you on the post-processing.

Carlos

[swellman]

Originally Posted by swellman

Reserving 1847_71_minus1.

A nice triple.

prp54 factor: 751675369654905088678231667221559046734934416341718309
prp60 factor: 860344312225890325626692969152005944727762796494291088905929
prp113 factor: 97939282732986918883657823741022254384311180414447883360700495089044693365385695961569969049390093405927645342427

TD=98(!) as attempts at 114 and 106 failed to build a matrix.

[Dubslow]

It shouldn't be hard for the server to track the original rel count before we download only uniques. Thanks for the tip Mike, I'll probably use that myself.