Houdini 1.03 is available

[sirabc]

On another board, someone tested 1.03 with a Intel Xeon X5355 Octa (8x2660 MHz). The node count was around 3.3 mn/s. Ivanhoe is around 7 mn/s. I've asked the tester whether Houdini was playing with only 4 cores, will post an update.

[Robert Houdart]

Hi Robert,
thanks for the new version.
I have started 2 separate matches on 2 different computers.So far I have already had 3 exceptions caused by Houdini.2 exceptions were on AMD 6 cores and one on Q9550.
Previous version was very stable but this new one for some reason is having problems.
Did anybody else so far had the similar problem with this new version.
robbolito

I haven't seen a single crash so far with Houdini 1.03, but obviously every change entails a potential risk. I'll run another stress test with 4 cores tonight.

By the way, I'm interested in the best value for Split_Depth on your 2 systems, could you try to determine the best value according to this explanation: Tuning the Split_Depth parameter on 4-core and 8-core hardware ?

With 4 cores or more the standard Split_Depth value of 10 may not be very well adapted to your hardware, in fact this might be a reason for Houdini's apparent lack of performance on multi-core AMD processors.

Robert

[sirabc]

1.03 with a Intel Xeon X5355 Octa (8x2660 MHz)

Update: Task manager reported 8 cores at 100%. Something maybe wrong with the 8 core version.

[Robert Houdart]

What are the results with the 4_CPU version?

For comparison, on a Core i5-750 (4 cores @ 2.66 GHz) I typically obtain around 7,000 kN/s in midgame positions, up to 11,000 kN/s in endgames.
If you're only seeing 3,300 kN/s you're obtaining the performance of less than 2 cores!

Clearly something is very wrong here.
A few suspects:
1) split_depth too low: you can test whether larger split_depth value solves anything
2) spinlocks for hash table don't work well on this hardware: I can compile a test version that doesn't use spinlocks

I'm ready to spend some time on this, but I'll need some reliable feedback. It would help if the person with the 8 core machine could join the discussion.

Robert

[Taner Altinsoy]

Thank you for this update. I have a few questions.
1) Is there a way to check whether large pages are working or not eg. like in rybka case checking the log file?
2) Do we need certaion amount of RAM (like 4GB) or hashtable size in order to be able to use large pages feature?
3) Does Split_Depth parameter also work for 2 core processors. Being an unlucky user of AMD I would like to see an improvement on this front.

Again thank you for creating and providing updates to this great engine.

Taner

[Robert Houdart]

Taner,

Thank you for this update. I have a few questions.
1) Is there a way to check whether large pages are working or not eg. like in rybka case checking the log file?
2) Do we need certaion amount of RAM (like 4GB) or hashtable size in order to be able to use large pages feature?
3) Does Split_Depth parameter also work for 2 core processors. Being an unlucky user of AMD I would like to see an improvement on this front.

1) Yes, in the UCI output you'll find either "xxx MB Large Page Hash used" or "xxx MB Large Page unavailable, using standard pages". If you find neither of the two your account may not have sufficient privileges to use large pages.
2) Even with 256 MB you'll find about 5% speed improvement.
3) Yes, please try whether this changes anything on your system.

Robert

[Vael Jean-Paul]

This is one off my positions i always use to test a engine:

You will see that Houdini 1.02 is the fastest to find the solution!
But i have also marked the KN/s and used as reference at time 07m07s
And because i see results from HD 1.03 default 64bit do better as the popcnt version..i check this out!
The results tell me that HD 1.03 popcnt is a little slower then HD 1.02 popcnt..
But HD 1.03 default x64 is much faster!! (Both larger pages ON)
So there must be something wrong during compiling a popcnt version!
When i check that with the IvanHoe's ,the popcnt version are always faster!
I have tried The Split_Depth parameter and changed to 13 ..but i get much slower result.
Was with the HD 1.03 popcnt version..will try the same with default version.

MAT in 17!

FEN: q3r1k1/5p2/3p1Rp1/np1P2P1/p1pPrQB1/2P5/PP5K/5R2 w - - 0 31

Houdini 1.0 x64 2cores: Depth 30 in 05m17s +M17! Core i7 920 @3.87Ghz (19x204) HT OFF
Houdini Devel x64 4cores: Depth 29 in 04m56s +M18 Core i7 920 @3.80Ghz (19x200) HT OFF
Depth 32 in 07m19s +M18
Depth 33 in 09m40s +M17
Houdini Devel V2 x64 4cores: Depth 23 in 01m58s +M19 Core i7 920 @3.80Ghz (19x200) HT OFF
Depth 30 in 03m57s +M18
Depth 32 in 05m38s +M18
Depth 33 in 07m07s +M18 (16400KN/s)
Depth 34 in 10m58s +M17 (16970KN/s)
Houdini 1.02 i7 x64 4cores: Depth 22 in 01m52s +M18 Core i7 920 @3.80Ghz (19x200) HT OFF
Depth 29 in 03m26s +M17!!
Depth 33 in 07m07s +M17 (16664KN/s)
Houdini 1.03 i7 x64 4cores: Depth 24 in 04m14s +M30 Core i7 920 @3.80Ghz (19x200) HT OFF
Depth 25 in 04m34s +M26
Depth 27 in 04m47s +M27
Depth 30 in 06m02s +M26 (07m07s -> 16200Kn/s)
Houdini 1.03 x64 4cores: Depth 23 in 01m39s +M43 Core i7 920 @3.80Ghz (19x200) HT OFF
Depth 24 in 01m51s +M19
Depth 29 in 02m48s +M18
Depth 32 in 04m28s +M18
Depth 33 in 06m02s +M17! (17551Kn/s) (07m07s -> 17770Kn/s)
Depth 34 in 07m33s +M17 (08m52s -> 18000Kn/s)

Hope this info is a little usefull.

JP.

[Robert Houdart]

Jean-Paul,

Thanks for the data, quite interesting.
Multi-core speed comparisons are not very straightforward: because of the random threading behaviour the different runs may actually be looking at different variations with different branching factors.

But I agree with your conclusion about the POPCNT versions, the speed difference is small and not necessarily in favor of the POPCNT version. The reason for this is that the Houdini code contains very few occurrences of the POPCNT instruction, I've tried to avoid this expensive function as much as possible.

In future Houdini releases I will probably no longer bother with separate POPCNT versions, the speed gain is too small to justify a separate compile.

Robert

[Vael Jean-Paul]

With the default version i get a better result with Split Depth 13!
I think i gonne run my test with these settings!

Houdini 1.03 x64 4cores: Depth 22 in 01m19s +M19 Core i7 920 @3.80Ghz (19x200) HT OFF (Split Depth = 13!) ->SD = 15 & 12 are slower!)
Depth 29 in 02m24s +M18
Depth 31 in 03m19s +M18 (16880KN/s)
Depth 32 in 04m04s +M18 (17250KN/s)
Depth 33 in 05m12s +M18 (17700KN/s) (07m07s -> 18364KN/s)
Depth 34 in 07m52s +M17 (18532KN/s)

Same run with SD=13 Depth 22 in 02m19s +M18
Depth 31 in 04m16s +M17! (06m37s -> 17487KN/s)

JP.

[artiom]

Hello Robert,

Thanks for your great job

I have a question - I use 32x Windows and 2 CPU and sometimes Houdine 1.03 just frozen!(no lines, nothing!) Say, after move 17 (I mean analysis option - 3 lines). And it's not only my review. Do we do something wrong or its just a bug?

Thanks again for your time!