Crafty 23.1 scaling problem on Nehalem octa

[Hugo]

Hello Gian-Carlo

core temp is arround 54C° at 100% load.

kind regards, Clemens

[Hugo]

Hello Mr. Hyatt

I think , the hardware is okay. I was asking Mr. Feist from ChessBase to build me a CB native compile of this new crafty engine. The result was the same low KNs on the the Nehalem octa machine AND the same bad scling now also on the skulltrail octa machine, which was doing well with the JA compile.
That is why I think it might be something inside the code, and not something hardware related.

Regards, Clemens Keck

The background, for all these questions is, that I am using crafty 23.1 in an online tournament. Right now, I am unsure if all works best on octa. Thats why I use a quad core machine for the tournament. I have to play 2 games (60min +15sec) per week. I would like to give crafty best hardware I have.
Also the cluster version would be verry interesting, but it think it is private.
I have 4 computers with infiniband cards. So I can build a high performant subnet for clustering (MPI or whatever).

[zullil]

Hello Mr. Hyatt

I think , the hardware is okay. I was asking Mr. Feist from ChessBase to build me a CB native compile of this new crafty engine. The result was the same low KNs on the the Nehalem octa machine AND the same bad scling now also on the skulltrail octa machine, which was doing well with the JA compile.
That is why I think it might be something inside the code, and not something hardware related.

Regards, Clemens Keck

Hi Clemens,

I compiled 23.1 and it seems to run very well on my dual quad-core Mac Pro. Now that I've discovered how to disable hyperthreading on my machine, I get 25M nps (from a pair of E5520's at 2.26 GHz).

I wonder if it's OS-related?

Louis

[Hugo]

hello Louis

I dont think so. I have a skulltrail octa ( two QX 9775's, 3,2 GHz) with the same OS. I get 22.000KNs on that machine.

Regards Clemens Keck

[bob]

Maybe the box is overheating and throttling itself when all CPUs are enabled?

So far as I know, from testing here, it only throttles back to nominal CPU clock speed as printed on the chip. However, that was one reason to disable the turbo-boost stuff anyway. It doesn't help when all cores are running, and will make a single or dual-core run go faster than expected. But to get down to 9M nodes per second on that box says something else is badly wrong.

[bob]

hello Louis

I dont think so. I have a skulltrail octa ( two QX 9775's, 3,2 GHz) with the same OS. I get 22.000KNs on that machine.

Regards Clemens Keck

Strange results, to be sure. It ran just fine on the i7 we had here for evaluation. I just ran it on one node of our dual quad-core i7 cluster and it runs as expected. The scaling is never exactly 8x, but it is usually never less than 7.5x either...

If the chessbase compile is running slow, and it now runs slow on both machines, it is likely something they have done, either with a bad compiler, bad optimization options, or bad changes they made since they do modify the Crafty source to make it work with their API...

[Hugo]

Hello Mr. Hyatt

it seems "we" cannot find the reasons yet.
I will use my quadcore machine for the tournament, using Jim Abletts's 64 bit compile. I am happy to see that crafty did a big jump in playing strenth.

Kind regards, Clemens Keck

[beachknight]

Hi,

Have you tried Jim's new x64 compile with pthread at:

http://www.mediafire.com/download.php?mdzynyl1onm

Best,

[Hugo]

Hello Harun,

yes I tried, but did not change anything on my machine.
thanx

Clemens

[beachknight]

Hello Harun,

yes I tried, but did not change anything on my machine.
thanx

Clemens

No problem. I am interested in this topic because
I am thinking of buying an octa in few months.
I hope to see crafty running with 25-30 mN on it.

Which octa system is the cheapest for my purposes?
Intel-based or AMD-based or another? Any advice?

TIA,

Best,