Msieve 1.50 feedback - Page 4

xilman · 2012-04-18, 08:45

Quote:

Originally Posted by xilman

Now to start experimenting.

Hmm. Not impressed so far:

Code:

linear algebra at 0.0%, ETA 17h 0m564656 dimensions (0.0%, ETA 17h 0m)    
checkpointing every 200000 dimensions656 dimensions (0.1%, ETA 18h38m)
...
linear algebra completed 27564 of 3564656 dimensions (0.8%, ETA 17h38m)
...
linear algebra completed 34649 of 3564656 dimensions (1.0%, ETA 17h44m)

No significant difference, AFAICT. That is the output from Tom's suggested command line, suitably augmented with msieve's requirements.

Oh well. Until I get a matrix too big to run comfortably on the single machine with -t 8 and building a cluster out of the boxes here becomes necessary, I'll keep the MPI version in reserve.

Thanks for your help and guidance guys.

Paul

fivemack · 2012-04-18, 23:11

Is -nc2 4,2 or taskset -c 0,1,4,5 mpirun -p 4 .... -nc2 2,2 .... any better?

(idea of the first is to see if it gets the decomposition of the matrix to fit the layout of the processors better; idea of the second is to use two jobs on each memory controller)

xilman · 2012-04-19, 12:58

Quote:

Originally Posted by xilman

Thought I'd update my systems to the latest msieve version from a variety of older ones. Since the last iteration of this exercise one system has had OpenBSD 5.0 installed. Msieve doesn't build from the standard tarball on this machine.

Well, I've been having fun. After much hacking of BSDmakefiles everything appears to build with the exception of the lattice sievers. This is not in the least bit surprising

I've been fighting Jens' code for many years now and am much too familiar with that particular tarpit.

When the next matrix becomes available here, I'll see what happens with the new build. Until then the machine can continue to run ecm clients in its spare time.

Paul

xilman · 2012-04-19, 13:52

Quote:

Originally Posted by xilman

Well, I've been having fun. After much hacking of BSDmakefiles everything appears to build with the exception of the lattice sievers. This is not in the least bit surprising

I've been fighting Jens' code for many years now and am much too familiar with that particular tarpit.

When the next matrix becomes available here, I'll see what happens with the new build. Until then the machine can continue to run ecm clients in its spare time.

Paul

Wow, I got it working with less hassle than expected. The 14e siever is running generic code and is only half the speed expected from 64-bit assembly but I'm amazed that it works at all!

fivemack · 2012-04-20, 09:27

Quote:

Originally Posted by xilman

I've a matrix which is 3564431 x 3564656 x 251590133 (70.58/col). Installed is stock OpenMPI on a Fedora 15 system.

Running "msieve -nc2 -t 8 ..." on a dual 4-core machine gives a predicted time of 15h38m --- in line with previous experience with other matrices of comparable size.

Using "mpirun -np 8 msieve -nc2 2x4 ..." on the same machine predicts 17h34m --- quite significantly longer.

Again I'm sorry if I'm teaching you to suck eggs, but which dual 4-core machine is this? I would only expect an improvement from MPI on a NUMA machine (Opteron or Xeon 5[56]xx or Xeon-E5-2xxx), I would expect a slowdown on a Xeon 5[1-4]xx.

xilman · 2012-04-20, 18:17

Quote:

Originally Posted by fivemack

Again I'm sorry if I'm teaching you to suck eggs, but which dual 4-core machine is this? I would only expect an improvement from MPI on a NUMA machine (Opteron or Xeon 5[56]xx or Xeon-E5-2xxx), I would expect a slowdown on a Xeon 5[1-4]xx.

Don't worry about egg-sucking matters. I barely know what I'm doing most of the time anyway. As you can see from the following snippet I'm on a 5[1-4]xx machine for which you would expect a slow-down.

Code:

[pcl@maat work]$ cat /proc/cpuinfo
processor	: 0
vendor_id	: GenuineIntel
cpu family	: 6
model		: 23
model name	: Intel(R) Xeon(R) CPU           E5440  @ 2.83GHz
stepping	: 10
microcode	: 0xa0b
cpu MHz		: 2826.047
cache size	: 6144 KB
physical id	: 0
siblings	: 4
core id		: 0
cpu cores	: 4
apicid		: 0
initial apicid	: 0
fpu		: yes
fpu_exception	: yes
cpuid level	: 13
wp		: yes
flags		: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx lm constant_tsc arch_perfmon pebs bts rep_good nopl aperfmperf pni dtes64 monitor ds_cpl vmx est tm2 ssse3 cx16 xtpr pdcm dca sse4_1 xsave lahf_lm dts tpr_shadow vnmi flexpriority
bogomips	: 5652.09
clflush size	: 64
cache_alignment	: 64
address sizes	: 38 bits physical, 48 bits virtual
power management:
...

The corresponding material from the 8-core machine I purchase from you is

Code:

-bash-4.2$ dmesg
OpenBSD 5.0 (GENERIC.MP) #63: Wed Aug 17 10:14:30 MDT 2011
    deraadt@amd64.openbsd.org:/usr/src/sys/arch/amd64/compile/GENERIC.MP
real mem = 34358624256 (32766MB)
avail mem = 33429889024 (31881MB)
...
cpu0 at mainbus0: apid 0 (boot processor)
cpu0: Quad-Core AMD Opteron(tm) Processor 2380, 2500.45 MHz
cpu0: FPU,VME,DE,PSE,TSC,MSR,PAE,MCE,CX8,APIC,SEP,MTRR,PGE,MCA,CMOV,PAT,PSE36,CFLUSH,MMX,FXSR,SSE,SSE2,HTT,SSE3,MWAIT,CX16,POPCNT,NXE,MMXX,F
FXSR,LONG,3DNOW2,3DNOW
cpu0: 64KB 64b/line 2-way I-cache, 64KB 64b/line 2-way D-cache, 512KB 64b/line 16-way L2 cache
cpu0: ITLB 32 4KB entries fully associative, 16 4MB entries fully associative
cpu0: DTLB 48 4KB entries fully associative, 48 4MB entries fully associative
cpu0: apic clock running at 200MHz

That machine now has msieve installed, though it took a bit of hassle and the installation of gmake to get it working. As yet I've not tried installing OpenMPI and rebuilding msieve on that system.

Paul

xilman · 2012-04-24, 19:04

Quote:

Originally Posted by xilman

Wow, I got it working with less hassle than expected. The 14e siever is running generic code and is only half the speed expected from 64-bit assembly but I'm amazed that it works at all!

Making good progress. The Perlwhich comes by default doesn't support mutlithreaded apps but the "experimental" code in factMsieve.pl for multiple clients appears to work well.

The current limitation is shown by this snapshot:

Code:

read 10M relations
error -9 reading relation 14532534
error -9 reading relation 14568142
error -15 reading relation 14788004
error -15 reading relation 14835861
error -15 reading relation 14873415
error -11 reading relation 14901514
read 20M relations
skipped 18 relations with b > 2^32
found 5121367 hash collisions in 26778492 relations
added 579 free relations
commencing duplicate removal, pass 2
found 5134747 duplicates and 21644324 unique relations
memory use: 165.2 MB
reading ideals above 720000
commencing singleton removal, initial pass
failed to reallocate 655360000 bytes

[1]+  Exit 255                ../msieve -s w7_233.dat -l w7_233.log -i w7_233.ini -t 8 -nf w7_233.fb -nc -v
-bash-4.2$

My best guess is that ulimits needs to be set appropriately.

With luck I'll be able to give concrete guidance on how to run msieve under BSD Real Soon Now. Given that MacOS is just a bastardized version of BSD the results may be of reasonably wide interest.

Paul

Batalov · 2012-04-24, 19:18

I can fire up my old SGI and post an (equally useful) recipe for sieving on IRIX! :-)
Next stop, DEC Alpha Tru64 Unix.

xilman · 2012-04-24, 19:27

Quote:

Originally Posted by xilman

My best guess is that ulimits needs to be set appropriately.

With luck I'll be able to give concrete guidance on how to run msieve under BSD Real Soon Now. Given that MacOS is just a bastardized version of BSD the results may be of reasonably wide interest.

Paul

Yup, /etc/login.conf had maximum memory set at 512M. Reset it to 32G and trying again.

xilman · 2012-04-24, 19:32

Quote:

Originally Posted by Batalov

I can fire up my old SGI and post an (equally useful) recipe for sieving on IRIX! :-)
Next stop, DEC Alpha Tru64 Unix.

My SillyGraphics box is in the loft, fairly close to the Alpha machine.

BSD is very much alive and kicking, unlike the other two. It's a good solid x86 OS, in boith 32 and 64 bit versions and, as I mentioned, extremely widely deployed under the name "MacOS".

In recent months I've become so disillusioned with the flakiness of recent Linux releases and with Gnome 3.x that I've returned to BSD for reliable systems. I first met BSD in 1981 or so and have never been disappointed by it other than the relative paucity of ports compared with Windoze and Linsux.

fivemack · 2012-04-24, 22:53

I've run msieve out of the box on a couple of Macs, and got reasonably good timings (four hours for a 2.2M matrix, 17.5 hours for 4.6M) just running -t8 - the much more difficult thing on Mac is getting the sievers compiled. I routinely use msieve for polynomial selection on up to quite large (170-digit) numbers on the dual-quad-xeon Mac Pro on my desk at work.

2012-04-24, 19:18	#41
Batalov "Serge" Mar 2008 Phi(4,2^7658614+1)/2 22431₈ Posts	I can fire up my old SGI and post an (equally useful) recipe for sieving on IRIX! :-) Next stop, DEC Alpha Tru64 Unix. Last fiddled with by Batalov on 2012-04-24 at 19:20

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Msieve 1.53 feedback	xilman	Msieve	149	2018-11-12 06:37
Msieve v1.48 feedback	Jeff Gilchrist	Msieve	48	2011-06-10 18:18
Msieve 1.43 feedback	Jeff Gilchrist	Msieve	47	2009-11-24 15:53
Msieve 1.42 feedback	Andi47	Msieve	167	2009-10-18 19:37
Msieve 1.41 Feedback	Batalov	Msieve	130	2009-06-09 16:01

2012-04-18, 23:11	#35
fivemack (loop (#_fork)) Feb 2006 Cambridge, England 2³×11×73 Posts	Is -nc2 4,2 or taskset -c 0,1,4,5 mpirun -p 4 .... -nc2 2,2 .... any better? (idea of the first is to see if it gets the decomposition of the matrix to fit the layout of the processors better; idea of the second is to use two jobs on each memory controller)

2012-04-24, 22:53	#44
fivemack (loop (#_fork)) Feb 2006 Cambridge, England 1918₁₆ Posts	I've run msieve out of the box on a couple of Macs, and got reasonably good timings (four hours for a 2.2M matrix, 17.5 hours for 4.6M) just running -t8 - the much more difficult thing on Mac is getting the sievers compiled. I routinely use msieve for polynomial selection on up to quite large (170-digit) numbers on the dual-quad-xeon Mac Pro on my desk at work.