mersenneforum.org  

Go Back   mersenneforum.org > Factoring Projects > NFS@Home

Reply
 
Thread Tools
Old 2018-03-31, 12:00   #1
fivemack
(loop (#_fork))
 
fivemack's Avatar
 
Feb 2006
Cambridge, England

13·491 Posts
Default Some numbers

This is for the same job (a 22.78M density-146 from a GNFS187); indeed, resuming from the same point.

Timings on four different reasonably chunky computers. Note how well the single-socket SKL is doing.

Haven't tried MPI on the dual-socket machines yet.

Speeds are in millions of dimensions per day.

Code:
		#cores	core	GHz	speed	RAM
oak		20	skl	2.2	2.669	EDDR4 2400 x6
butternut	6	hsw	3.3	1.427	DDR4 2400 x4 
pineapple	14	skl	3.1	3.454	DDR4 2666 x4
birch4		16	snb	2	2.003	EDDR3 1066 x4

Last fiddled with by fivemack on 2018-04-09 at 13:32
fivemack is offline   Reply With Quote
Old 2018-04-06, 06:47   #2
frmky
 
frmky's Avatar
 
Jul 2003
So Cal

2×3×347 Posts
Default

SKL is great. Put one msieve process per socket and use threads to distribute to the cores.
frmky is online now   Reply With Quote
Old 2018-04-06, 08:59   #3
pinhodecarlos
 
pinhodecarlos's Avatar
 
"Carlos Pinho"
Oct 2011
Milton Keynes, UK

2·7·349 Posts
Default

Greg,

Looking at the status page for NFS, example, one SNFS 311. What would be the machine specs to do the post-processing in two weeks time.

Make me a list please: dual socket, amount of memory, etc.
pinhodecarlos is online now   Reply With Quote
Old 2018-04-13, 09:39   #4
fivemack
(loop (#_fork))
 
fivemack's Avatar
 
Feb 2006
Cambridge, England

13·491 Posts
Default

Quote:
Originally Posted by pinhodecarlos View Post
Greg,

Looking at the status page for NFS, example, one SNFS 311. What would be the machine specs to do the post-processing in two weeks time.

Make me a list please: dual socket, amount of memory, etc.
Doing post-processing for an SNFS 311 is a large job; https://pastebin.com/cXqsPFDL indicates that 32 computers each of which was a dual-socket 12-core Skylake Xeon was used, though (if the machines had been devoted 24/365) it would have done the matrix in about 3.5 days rather than about two weeks.

So you could probably get away with an MPI grid of eight dual-socket 12-core SKL Xeons; call it £40,000 (you don't need much memory in each node, but you probably want a fast interconnect; a 12-port 40Gb Infiniband switch is $3000 and the adapters are $500 per node)

The filtering would fit on a 64GB machine; using the normal n^2 scaling, my SKL which is taking 530 hours for a 37M matrix would take about two months. That's about a £3000 machine (mine was a bit more because it has a GTX1080Ti in it).
fivemack is offline   Reply With Quote
Old 2018-04-13, 09:58   #5
pinhodecarlos
 
pinhodecarlos's Avatar
 
"Carlos Pinho"
Oct 2011
Milton Keynes, UK

131616 Posts
Default

When we meet again can we discuss this in more deep please. I’ll send an invitation in due course since I am with a high work load and also I’m waiting for the Bristish summer.
pinhodecarlos is online now   Reply With Quote
Old 2018-04-13, 14:05   #6
xilman
Bamboozled!
 
xilman's Avatar
 
"𒉺𒌌𒇷𒆷𒀭"
May 2003
Down not across

22·3·887 Posts
Default

Quote:
Originally Posted by pinhodecarlos View Post
and also I’m waiting for the Bristish summer.
I believe that it is scheduled for August 21st and 22nd this year.
xilman is offline   Reply With Quote
Old 2018-04-22, 13:37   #7
fivemack
(loop (#_fork))
 
fivemack's Avatar
 
Feb 2006
Cambridge, England

11000111011112 Posts
Default Some MPI numbers

These are times between 'commencing Lanczos iteration' and 'recovered non-trivial dependencies' for a 1478190 x 1478415 matrix at density 70 (from the C135 from term 5143 of aliquot sequence 219240)

Code:
name    MPI  cores   seconds
oak     1x2  10 x2   1167
oak     1x1  10      1302
oak     1x1  10+10HT 1408
oak     1x1  20      1212
birch1  1x1  16+16HT 1655
birches 1x8  8 x8    1285
birches 2x4  8 x8    937
birches 4x2  8 x8    960
birches 8x1  8 x8    1419
oak is a dual-socket Xeon Silver 4114; the birches are four (ex-Facebook-I-think) dual-socket E5-2650 Sandy Bridge machines, attached by fibre to a 10Gbit network switch.

On oak I launch the MPI job with
Code:
mpirun --report-bindings -np 2 --map-by socket:PE=10 --bind-to core /home/nfsworld/msieve-svn/MPI/msieve -v -nc2 -t 10
For the birches the mpirun is
Code:
for u in "1,8" "2,4" "4,2" "8,1"; do mpirun --report-bindings -np 8 -hostfile mpi-hosts /home/nfsworld/msieve-svn/MPI/msieve -v -nc2 $u -t 8; done
and the MPI host list is
Code:
birch@birch1.fivemack.internal slots=2
birch@birch2.fivemack.internal slots=2
birch@birch3.fivemack.internal slots=2
birch@birch4.fivemack.internal slots=2
While running the job I see about 340MB/sec read and write traffic reported on the network interface of each birch.

I don't see CPU occupancy much above 30% for any of the cores on birch, or much above 60% on oak.

Last fiddled with by fivemack on 2018-09-29 at 06:08
fivemack is offline   Reply With Quote
Old 2018-04-22, 13:41   #8
fivemack
(loop (#_fork))
 
fivemack's Avatar
 
Feb 2006
Cambridge, England

11000111011112 Posts
Default

Quote:
Originally Posted by fivemack View Post
Doing post-processing for an SNFS 311 is a large job; https://pastebin.com/cXqsPFDL indicates that 32 computers each of which was a dual-socket 12-core Skylake Xeon was used, though (if the machines had been devoted 24/365) it would have done the matrix in about 3.5 days rather than about two weeks.
I made a mistake parsing those results; there are probably 16 computers each of which is a dual 24-core Skylake Xeon (unless it is a very fancy cluster with 4 eight-socket 24-core machines), so that's $130k for the processors alone.

Last fiddled with by fivemack on 2018-04-22 at 13:41
fivemack is offline   Reply With Quote
Old 2018-04-22, 14:38   #9
VictordeHolland
 
VictordeHolland's Avatar
 
"Victor de Hollander"
Aug 2011
the Netherlands

23·3·72 Posts
Default

So the take away from this is? Don't use MPI for LA unless you need results quickly?
VictordeHolland is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
Carmichael numbers and Devaraj numbers devarajkandadai Number Theory Discussion Group 0 2017-07-09 05:07
6 digit numbers and the mersenne numbers henryzz Math 2 2008-04-29 02:05
LLT numbers, linkd with Mersenne and Fermat numbers T.Rex Math 4 2005-05-07 08:25

All times are UTC. The time now is 19:29.

Wed Apr 14 19:29:05 UTC 2021 up 6 days, 14:09, 0 users, load averages: 1.85, 2.17, 2.08

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

This forum has received and complied with 0 (zero) government requests for information.

Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation.
A copy of the license is included in the FAQ.