2x Xeon x5450 QuadCore Workstation for factorization?

[beans]

Hi,

I want to buy a computer and want to use it for large integer factorization. I have the opportunity to get an old FSC Celsius R550 workstation with a D2569 Dual Socket 771 board and 2 Xeon 5450 for about 500€. Is it useful to invest in such a machine ore will do a single Intel i7 Quad a better job on gnfs-lasieve?

Thanks.

[fivemack]

That's an interesting decision point.

Sieving the same region with the same polynomial on a Xeon 5420 and an i7-2600S I get

(0.10065 sec/rel) (Xeon 5420, one thread) - note this is 2.5GHz rather than the 3.0GHz of a 5450
(0.05327 sec/rel) (i7-2600S, one turbo-boosted thread)
(0.06987 sec/rel) (i7-2600S, four threads running on one core each, time for median thread)
(0.11194 sec/rel) (i7-2600S, eight threads running on eight hypercores, time for median thread)

So the Xeon workstation may be a bit faster in aggregate (though note that the i7/2600S is a low-power model at 2.8GHz whilst you'd probably be using a 3770K at 3.5GHz), at the price of using a *lot* more electricity. I would get a minimal Ivy Bridge box instead - I think that can be done for 500€

[henryzz]

Not mentioning that you would have to handle twice as many cores.

[fivemack]

(my colleague turns out to be out of the office today so I was able to run the eight-threads test)

[fivemack]

Originally Posted by henryzz

Not mentioning that you would have to handle twice as many cores.

No, for GNFS sieving you want to run eight threads on either a dual quad-Core2-Xeon or a single quad-{Sandy,Ivy}Bridge with hyperthreading.

[henryzz]

Originally Posted by fivemack

No, for GNFS sieving you want to run eight threads on either a dual quad-Core2-Xeon or a single quad-{Sandy,Ivy}Bridge with hyperthreading.

How much gain does hyperthreading gain you?

[fivemack]

Originally Posted by henryzz

How much gain does hyperthreading gain you?

(0.06987 sec/rel) (i7-2600S, four threads running on one core each, time for median thread)
(0.11194 sec/rel) (i7-2600S, eight threads running on eight hypercores, time for median thread)

so with four threads you get 57.2 relations per second and with eight you get 71.4; so hyperthreading is gaining about 25%

You know, you could have done that arithmetic yourself.

[henryzz]

Originally Posted by fivemack

(0.06987 sec/rel) (i7-2600S, four threads running on one core each, time for median thread)
(0.11194 sec/rel) (i7-2600S, eight threads running on eight hypercores, time for median thread)

so with four threads you get 57.2 relations per second and with eight you get 71.4; so hyperthreading is gaining about 25%

You know, you could have done that arithmetic yourself.

Sorry I didn't see the numbers above when I made that post.

[fivemack]

On an i7/960 at the office (so hyper-threading but not much turboing)

one thread 0.06429s 15.6r/s
six threads 0.06853s 87.6r/s
twelve threads 0.10810s 111r/s

so here I'm getting a 27% boost from HT, about the same as the SNB.

(the Q6600 downstairs gets me 0.10033s/r, one thread on the 48-thread K10 gets 0.12355)

The test is /home/nfsworld/gnfs-batalov-old/gnfs-lasieve4I14e -a testcase.txt -f 16000000 -c 2000 with the testcase.txt file attached; I'd be interested to see figures from people with Sandy/Ivy Bridges, and particularly I'd be interested in seeing N-thread and 2N-thread runs on Bulldozers. 64-bit Linux results would be comparable with the ones I've given here.

[Dubslow]

Why would HT help? In theory, one full core should be able to do the same work as two HThreads on one core.

[fivemack]

Originally Posted by Dubslow

Why would HT help? In theory, one full core should be able to do the same work as two HThreads on one core.

That's not true in theory, which is why it's not true in practice. It's not a matter of alternating issuing a line of micro-ops from each thread; what happens is that each thread pushes micro-ops into a separate queue, and the issue unit can pull micro-ops out of either queue to fill the execution units.

If the code were sufficiently superbly optimised that every execution unit could be filled at every cycle by either thread, then this would come out the same as alternation; but there's very little code (though prime95 is probably an example that comes close) which is that optimised. Generally there are opportunities for things from thread 2 to fill holes in thread 1 and vice versa.