i7-3930K benchmarks

James Heinrich · 2011-12-17, 01:55

Quote:

Originally Posted by fivemack

I see you have a 3930K now. Would you be prepared to do what was done at http://www.mersenneforum.org/showpos...01&postcount=7
(though -t 3, -t 6, -t 12 are more sensible thread-counts to test on a six-core)
and also running msieve149.exe from msieve149.zip downloaded from
http://sourceforge.net/projects/msie...sieve%20v1.49/
Please post the answers in the msieve forum http://www.mersenneforum.org/forumdisplay.php?f=83
I am extremely interested in how fast the four-channel memory controller is for msieve linear algebra as the thread count increases; if it is interestingly fast then I'll probably get a 3930K next month.

I'm just posting here at fivemack's request, I don't actually understand what these results mean.

v1.39 doesn't seem to thread very well. Even with -t 12, at best I saw 6 physical cores in use, with each core running at about 60% capacity.

v1.49 seems to thread much better, at -t 12 it's spread evenly across 12 hyperthreaded cores at about 65-70% load evenly on all cores.

Attached log file includes (in order):
1.39: msieve.exe -v -nc -t 12
1.39: msieve.exe -v -nc -t 6
1.49: msieve.exe -v -nc -t 12
1.49: msieve.exe -v -nc -t 6
1.49: msieve.exe -v -nc -t 3

fivemack · 2011-12-17, 07:10

Something odd is going on here: 1.39 is faster than 1.49 even though it's reporting lower CPU usage and (since filtering has been improved between 1.39 and 1.49) running a larger matrix.

$ \begin{tabular}{|l|l|l|l|l|l|} \hline & time (3)& time (6)& time (12) \\ \hline \\ 1.39& & 1245 & 1301 \\ 1.49& 2363 & 1653 & 1571 \\ \hline \end{tabular} $

pinhodecarlos · 2011-12-17, 09:40

Could you also post the spec of your machine? (memory speed, motherboard etc)

Thank you in advance.

James Heinrich · 2011-12-17, 14:04

Quote:

Originally Posted by pinhodecarlos

Could you also post the spec of your machine? (memory speed, motherboard etc)

Sorry, yes:

CPU: Intel i7-3930K @ 4500MHz
RAM: Corsair Vengeance 4*8GB [quad-channel] (CMZ32GX3M4X1600C10) DDR3-1600 10-10-10-27-2T
motherboard: Asus P9X79 PRO

Let me know if I missed anything.

em99010pepe · 2011-12-17, 15:53

Core i5 750@3.6GHz, 8 GB DDR3 1333 MHz.

Code:

Sat Dec 17 15:22:31 2011  
Sat Dec 17 15:22:31 2011  
Sat Dec 17 15:22:31 2011  Msieve v. 1.50 (SVN unknown)
Sat Dec 17 15:22:31 2011  random seeds: df0f8cb4 76141666
Sat Dec 17 15:22:31 2011  factoring 4173963189037240309713561470980060844301599530724369583171546744398445920275944654104909360191816721597816377363188552009346572473 (130 digits)
Sat Dec 17 15:22:32 2011  searching for 15-digit factors
Sat Dec 17 15:22:32 2011  commencing number field sieve (130-digit input)
Sat Dec 17 15:22:32 2011  R0: -10678489767165051962024349
Sat Dec 17 15:22:32 2011  R1: 202148791722581
Sat Dec 17 15:22:32 2011  A0: 130068846150643433934522883496770
Sat Dec 17 15:22:32 2011  A1: 1272305145809368946430393037
Sat Dec 17 15:22:32 2011  A2: -4820655295278675156855
Sat Dec 17 15:22:32 2011  A3: -39469582197230476
Sat Dec 17 15:22:32 2011  A4: 38243516789
Sat Dec 17 15:22:32 2011  A5: 30060
Sat Dec 17 15:22:32 2011  skew 228570.75, size 1.569e-012, alpha -6.792, combined = 7.496e-011 rroots = 5
Sat Dec 17 15:22:32 2011  
Sat Dec 17 15:22:32 2011  commencing relation filtering
Sat Dec 17 15:22:32 2011  estimated available RAM is 8183.1 MB
Sat Dec 17 15:22:32 2011  commencing duplicate removal, pass 1
Sat Dec 17 15:24:32 2011  found 2598477 hash collisions in 13708853 relations
Sat Dec 17 15:24:50 2011  added 1 free relations
Sat Dec 17 15:24:50 2011  commencing duplicate removal, pass 2
Sat Dec 17 15:24:55 2011  found 2526833 duplicates and 11182021 unique relations
Sat Dec 17 15:24:55 2011  memory use: 82.6 MB
Sat Dec 17 15:24:55 2011  reading ideals above 720000
Sat Dec 17 15:24:56 2011  commencing singleton removal, initial pass
Sat Dec 17 15:26:47 2011  memory use: 344.5 MB
Sat Dec 17 15:26:47 2011  reading all ideals from disk
Sat Dec 17 15:26:47 2011  memory use: 330.4 MB
Sat Dec 17 15:26:48 2011  commencing in-memory singleton removal
Sat Dec 17 15:26:49 2011  begin with 11182021 relations and 11719970 unique ideals
Sat Dec 17 15:26:54 2011  reduce to 5210969 relations and 4863630 ideals in 16 passes
Sat Dec 17 15:26:54 2011  max relations containing the same ideal: 96
Sat Dec 17 15:26:56 2011  removing 825491 relations and 719218 ideals in 106273 cliques
Sat Dec 17 15:26:56 2011  commencing in-memory singleton removal
Sat Dec 17 15:26:56 2011  begin with 4385478 relations and 4863630 unique ideals
Sat Dec 17 15:26:58 2011  reduce to 4293130 relations and 4049611 ideals in 10 passes
Sat Dec 17 15:26:58 2011  max relations containing the same ideal: 83
Sat Dec 17 15:27:00 2011  removing 617219 relations and 510946 ideals in 106273 cliques
Sat Dec 17 15:27:00 2011  commencing in-memory singleton removal
Sat Dec 17 15:27:00 2011  begin with 3675911 relations and 4049611 unique ideals
Sat Dec 17 15:27:02 2011  reduce to 3609739 relations and 3470882 ideals in 10 passes
Sat Dec 17 15:27:02 2011  max relations containing the same ideal: 73
Sat Dec 17 15:27:03 2011  relations with 0 large ideals: 486
Sat Dec 17 15:27:03 2011  relations with 1 large ideals: 2707
Sat Dec 17 15:27:03 2011  relations with 2 large ideals: 35991
Sat Dec 17 15:27:03 2011  relations with 3 large ideals: 200230
Sat Dec 17 15:27:03 2011  relations with 4 large ideals: 592581
Sat Dec 17 15:27:03 2011  relations with 5 large ideals: 989462
Sat Dec 17 15:27:03 2011  relations with 6 large ideals: 978596
Sat Dec 17 15:27:03 2011  relations with 7+ large ideals: 809686
Sat Dec 17 15:27:03 2011  commencing 2-way merge
Sat Dec 17 15:27:05 2011  reduce to 2184638 relation sets and 2045781 unique ideals
Sat Dec 17 15:27:05 2011  commencing full merge
Sat Dec 17 15:27:26 2011  memory use: 228.9 MB
Sat Dec 17 15:27:27 2011  found 1094722 cycles, need 1073981
Sat Dec 17 15:27:27 2011  weight of 1073981 cycles is about 75457035 (70.26/cycle)
Sat Dec 17 15:27:27 2011  distribution of cycle lengths:
Sat Dec 17 15:27:27 2011  1 relations: 130527
Sat Dec 17 15:27:27 2011  2 relations: 127368
Sat Dec 17 15:27:27 2011  3 relations: 125795
Sat Dec 17 15:27:27 2011  4 relations: 113119
Sat Dec 17 15:27:27 2011  5 relations: 102993
Sat Dec 17 15:27:27 2011  6 relations: 87941
Sat Dec 17 15:27:27 2011  7 relations: 76759
Sat Dec 17 15:27:27 2011  8 relations: 65165
Sat Dec 17 15:27:27 2011  9 relations: 54780
Sat Dec 17 15:27:27 2011  10+ relations: 189534
Sat Dec 17 15:27:27 2011  heaviest cycle: 21 relations
Sat Dec 17 15:27:27 2011  commencing cycle optimization
Sat Dec 17 15:27:28 2011  start with 6170552 relations
Sat Dec 17 15:27:35 2011  pruned 146025 relations
Sat Dec 17 15:27:35 2011  memory use: 206.5 MB
Sat Dec 17 15:27:35 2011  distribution of cycle lengths:
Sat Dec 17 15:27:35 2011  1 relations: 130527
Sat Dec 17 15:27:35 2011  2 relations: 130418
Sat Dec 17 15:27:35 2011  3 relations: 130196
Sat Dec 17 15:27:35 2011  4 relations: 115818
Sat Dec 17 15:27:35 2011  5 relations: 105171
Sat Dec 17 15:27:35 2011  6 relations: 88903
Sat Dec 17 15:27:35 2011  7 relations: 77655
Sat Dec 17 15:27:35 2011  8 relations: 64752
Sat Dec 17 15:27:35 2011  9 relations: 54165
Sat Dec 17 15:27:35 2011  10+ relations: 176376
Sat Dec 17 15:27:35 2011  heaviest cycle: 20 relations
Sat Dec 17 15:27:36 2011  RelProcTime: 304
Sat Dec 17 15:27:36 2011  
Sat Dec 17 15:27:36 2011  commencing linear algebra
Sat Dec 17 15:27:36 2011  read 1073981 cycles
Sat Dec 17 15:27:37 2011  cycles contain 3464474 unique relations
Sat Dec 17 15:28:08 2011  read 3464474 relations
Sat Dec 17 15:28:11 2011  using 20 quadratic characters above 134217498
Sat Dec 17 15:28:25 2011  building initial matrix
Sat Dec 17 15:28:53 2011  memory use: 453.8 MB
Sat Dec 17 15:28:54 2011  read 1073981 cycles
Sat Dec 17 15:28:55 2011  matrix is 1073793 x 1073981 (326.8 MB) with weight 103906487 (96.75/col)
Sat Dec 17 15:28:55 2011  sparse part has weight 72792856 (67.78/col)
Sat Dec 17 15:29:01 2011  filtering completed in 2 passes
Sat Dec 17 15:29:02 2011  matrix is 1071372 x 1071560 (326.6 MB) with weight 103797093 (96.87/col)
Sat Dec 17 15:29:02 2011  sparse part has weight 72757804 (67.90/col)
Sat Dec 17 15:29:03 2011  matrix starts at (0, 0)
Sat Dec 17 15:29:04 2011  matrix is 1071372 x 1071560 (326.6 MB) with weight 103797093 (96.87/col)
Sat Dec 17 15:29:04 2011  sparse part has weight 72757804 (67.90/col)
Sat Dec 17 15:29:04 2011  saving the first 48 matrix rows for later
Sat Dec 17 15:29:04 2011  matrix includes 64 packed rows
Sat Dec 17 15:29:04 2011  matrix is 1071324 x 1071560 (314.1 MB) with weight 82570139 (77.06/col)
Sat Dec 17 15:29:04 2011  sparse part has weight 71622900 (66.84/col)
Sat Dec 17 15:29:04 2011  using block size 65536 for processor cache size 8192 kB
Sat Dec 17 15:29:08 2011  commencing Lanczos iteration (4 threads)
Sat Dec 17 15:29:08 2011  memory use: 268.3 MB
Sat Dec 17 15:29:13 2011  linear algebra at 0.1%, ETA 0h47m
Sat Dec 17 15:29:15 2011  checkpointing every 1210000 dimensions

em99010pepe · 2011-12-28, 14:24

Anyone here with more benches?

bsquared · 2011-12-29, 16:02

Here is some ETA based data for two different machines. I can confirm that 1.39 appears to be faster than 1.49, although it doesn't scale quite as well. Utilization with 8 cores on machine 1 using 1.39 hovered around 55%, while with 1.49 it was around 85%.

machine 1 (hyperthreading disabled):
Dual Intel(R)Xeon(R) CPU X5687 @ 3.60GHz (4 cores)
triple channel DDR3 @ 1333 MHz (per CPU) + QPI CPU interconnect

Code:

 
v1.49 (win 7 pro, 64 bit):
threads, ETA (min) (after settling)
2, 66
4, 41
8, 30
 
v1.39 (win 7 pro, 64 bit):
threads, ETA (min) (after settling)
2, 48
4, 32
8, 29

machine 2 (hyperthreading enabled):
Dual Intel(R)Xeon(R) CPU X5680 @ 3.33GHz (6 cores)
triple channel DDR3 @ 1333 MHz (per CPU) + QPI CPU interconnect

Code:

 
 
v1.50 (linux, 64 bit)
threads, ETA (min) (after settling)
4, 58
6, 45
8, 41
12, 39
24, 53

fivemack · 2011-12-29, 17:12

Do you have the ability to try out the MPI version on that dual-hex-Xeon machine? It was a distinct improvement on my (now xilman's) dual-quad-Opteron, see

http://www.mersenneforum.org/showpos...9&postcount=60
http://www.mersenneforum.org/showpos...9&postcount=65

bsquared · 2011-12-29, 18:38

Quote:

Originally Posted by fivemack

Do you have the ability to try out the MPI version on that dual-hex-Xeon machine? It was a distinct improvement on my (now xilman's) dual-quad-Opteron, see

http://www.mersenneforum.org/showpos...9&postcount=60
http://www.mersenneforum.org/showpos...9&postcount=65

As of now, no; mpi isn't installed and I don't have admin privs on the machine. Though our admin knows what I occasionally use these machines for and might be willing to install it for me. I'll check when he's back next week.

jasonp · 2011-12-29, 19:35

Other than adding MPI, the only other major change between v1.39 and v1.50 is that now vector-vector multiplies are handled by multiple threads. Even if that causes cache thrashing, it shouldn't cost 25% of the total performance...

Both MPI and multithreaded vector-vector code were added in v1.46. Can someone see if the regression happens between v1.45 and v1.46? Unfortunately the SF page only has v1.41 and up, I can provide older releases if necessary.

schickel · 2012-01-27, 08:55

Ooops....I thought I had put 1.49 on my hexcore to test, but looking at the log, I only had 1.39 on there, so I'll have to re-run the benches again. But just for yucks, here is a log for the LA on my box for 1.39.

Summary of run times for 1-6 threads with nothing else running:

Code:

elapsed time 02:56:55
elapsed time 01:39:01
elapsed time 01:15:45
elapsed time 01:05:25
elapsed time 01:00:09
elapsed time 00:59:16

I'll re-run the tests and then play around with seeing what else I can run on the system without impacting msieve too much.

Full specs on the box:

Gigabyte M68MT-S2P MOBO
8 GB RAM (2x-4GB DDR3 PC3-1066 1333 MHz Patriot)
Athlon Phenom II X6 1090T BE 3.2 GHz

2011-12-17, 07:10	#2
fivemack (loop (#_fork)) Feb 2006 Cambridge, England 3·2,141 Posts	Something odd is going on here: 1.39 is faster than 1.49 even though it's reporting lower CPU usage and (since filtering has been improved between 1.39 and 1.49) running a larger matrix. $<br /> \begin{tabular}{\|l\|l\|l\|l\|l\|l\|}<br /> \hline<br /> & time (3)& time (6)& time (12) \\ \hline \\<br /> 1.39& & 1245 & 1301 \\<br /> 1.49& 2363 & 1653 & 1571 \\<br /> \hline<br /> \end{tabular}<br />$ Last fiddled with by fivemack on 2011-12-17 at 09:54

2011-12-29, 16:02	#7
bsquared "Ben" Feb 2007 2²×3×293 Posts	Here is some ETA based data for two different machines. I can confirm that 1.39 appears to be faster than 1.49, although it doesn't scale quite as well. Utilization with 8 cores on machine 1 using 1.39 hovered around 55%, while with 1.49 it was around 85%. machine 1 (hyperthreading disabled): Dual Intel(R)Xeon(R) CPU X5687 @ 3.60GHz (4 cores) triple channel DDR3 @ 1333 MHz (per CPU) + QPI CPU interconnect Code: v1.49 (win 7 pro, 64 bit): threads, ETA (min) (after settling) 2, 66 4, 41 8, 30 v1.39 (win 7 pro, 64 bit): threads, ETA (min) (after settling) 2, 48 4, 32 8, 29 machine 2 (hyperthreading enabled): Dual Intel(R)Xeon(R) CPU X5680 @ 3.33GHz (6 cores) triple channel DDR3 @ 1333 MHz (per CPU) + QPI CPU interconnect Code: v1.50 (linux, 64 bit) threads, ETA (min) (after settling) 4, 58 6, 45 8, 41 12, 39 24, 53

2011-12-29, 19:35	#10
jasonp Tribal Bullet Oct 2004 DD7₁₆ Posts	Other than adding MPI, the only other major change between v1.39 and v1.50 is that now vector-vector multiplies are handled by multiple threads. Even if that causes cache thrashing, it shouldn't cost 25% of the total performance... Both MPI and multithreaded vector-vector code were added in v1.46. Can someone see if the regression happens between v1.45 and v1.46? Unfortunately the SF page only has v1.41 and up, I can provide older releases if necessary. Last fiddled with by jasonp on 2011-12-29 at 19:38

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Benchmarks	Pjetrode	Information & Answers	3	2018-01-07 23:23
Where are the Benchmarks	Sandman192	Homework Help	17	2012-04-05 19:03
Benchmarks	MurrayInfoSys	Information & Answers	3	2011-04-14 17:10
Benchmarks for 24.12	Prime95	Software	60	2005-06-11 07:35
Benchmarks	Vandy	Hardware	6	2002-10-28 13:45

2011-12-17, 09:40	#3
pinhodecarlos "Carlos Pinho" Oct 2011 Milton Keynes, UK 3×17×97 Posts	Could you also post the spec of your machine? (memory speed, motherboard etc) Thank you in advance.

2011-12-28, 14:24	#6
em99010pepe Sep 2004 2·5·283 Posts	Anyone here with more benches?

2011-12-29, 17:12	#8
fivemack (loop (#_fork)) Feb 2006 Cambridge, England 3×2,141 Posts	Do you have the ability to try out the MPI version on that dual-hex-Xeon machine? It was a distinct improvement on my (now xilman's) dual-quad-Opteron, see http://www.mersenneforum.org/showpos...9&postcount=60 http://www.mersenneforum.org/showpos...9&postcount=65