Not logged inRybka Chess Community Forum
Up Topic The Rybka Lounge / Computer Chess / chess engine speed/settings on my new 48c workstation
- - By jakall [de] Date 2017-12-03 20:07
Hello everybody, I need some advice about the chess engines on my new workstation, which I use for chess analysis / correspondence chess.
My machine is quite high-end, key specs are as follows: Dual Xeon Platinum 8168 (total of 48c/96t @ 2.7 GHz), 128 GB (8 x 16 GB) 2666 MHz DDR4 ECC RDIMM, 2 x 1 TB SSD, OS = Win 10 Pro for Workstations. (In MAXON's Cinebench R15 CPU benchmark it consistently scores around the 7000 mark.)

As chess engines I use Houdini 6.03 pro, Komodo 11.2.2 pro and the stockfish clone "asmFishW_2017-11-19_bmi2" on ChessBase 14 with the Syzygy tablebases.
I select 48 cores and (usually) use 32 GB as hash value for each of these engines.

Here are my 2 main issues:

1) I have been comparing my engine analysis data with what is shown in TCEC final. Despite my machine being nominally stronger than the TCEC computer (which is 44c/88t @ 2.2 GHz + 32 GB Ram), with Komodo 11.2.2 I cannot get to TCEC's level of nodes per second: my Komodo 11.2.2 initially (after hash has loaded, more on that later) achieves around 20000 kn/s and only after around 15 minutes uptime increases to 40000 kn/s. 44000 kn/s is the highest value I have seen with Komodo 11.2.2 on my machine, yet TCEC's Komodo gets to around 55000 - 65000 kn/s on a regular basis. Does anybody have an idea what could be the reason for this? A wrong engine setting on my part or should I maybe disable hyperthreading in my BIOS?
(With Houdini 6.03 on the other hand, I get quite a good 80000 - 90000 kn/s which is a bit higher than TCEC's Houdini 6.03 value. Actually, whenever I enter a new move in analysis mode, it initially even shows up to 130 Mn/s but quickly goes back down to 80000 - 90000 kn/s again. With asmFishW_2017-11-19_bmi2 I get around 80000 kn/s, which sounds reasonable although I am not quite sure how to evaluate that.)

2) Every time I activate an analysis engine it takes quite a long time until hash is "loaded". In case of 32 GB hash (my usual choice) it takes around 1 minute, for 64 GB around 2 minutes. Task manager - performance - RAM shows a graph that is steadily increasing by around 0.5 GB per second. Even before this process is completed, task manager shows that all 48 of 96 CPU threads are in use already, yet the selected engine only achieves values of 800 - 2000 kn/s (maybe that's the value for 1 core!?). After around 1-2 minutes when hash is loaded completely, engine speed immediately increases to their respective standard value of 20000 kn/s (in case of Komodo 11.2.2 within the first minutes of use) or 80000 - 90000 kn/s (in case of Houdini 6.03, asmFish).
Since I like to regularly switch between engines during my analysis, having to wait at least 1 minute each time for another engine loading is quite annoying. Also, I have to wait for around 10 seconds whenever I deavtivate an analysis engine while ChessBase shows the "unloading engine" window.
Any advice whether this is normal behaviour or, if not, what could be done to resolve that issue would be highly appreciated.

Thank you and best regards,
jakall

P.S.: Those are my further engine settings (as stated, hash selected = usually 32 GB, sometimes 64 GB):
Houdini 6.03 pro: threads 48, tactical mode 0, contempt 2, analysis contempt NO, EGTB Probe Depth 1, EGTB Fifty Moves Rule YES, NUMA offset 0, NUMA enabled YES, strength 100, never clear hash NO, hash file <empty>, fiftymovesdistance 50, move overhead 0, UCI log file (none), hide redundant output NO, own book NO, book file <empty>, best book line NO

asmFishW_2017-11-19_bmi2: LargePages YES, threads 48, nodeaffinity all, priority none, LogFile <empty>, TTFile <empty>, contempt 0, MoveOverhead 50, SyzygyProbeDepth 1, SyzygyProbeLimit 6, Syzygy50MoveRule YES, OwnBook NO, BookFile <empty>, BestBookMove NO, BookDepth 100

Komodo 11.2.2 64 bit pro: threads 48, table memory 1024, OwnBook YES, BookFile komodo.bin, BestBookLine NO, LogFile <empty>, Hash File Name <empty>, Use LMR YES, Use Null Move Pruning YES, Minimal Reporting 0, Overhead ms 30, Time Usage 0, Use Syzygy YES, Syzygy Probe Depth 2, Syzygy Probe Limit 6,  Syzygy 50 Move Rule YES, Smart Syzygy NO, King Safety 73, Selectivity 120, Reduction 0, Dynamism 130, Contempt 10, White Contempt NO, Null Scale 28, Progress Threshold 30, Variety 0
Parent - By Banned for Life (Gold) Date 2017-12-03 20:29
Lukas Cimiotti would be your best source for advice on these issues...
Parent - By user923005 (****) [us] Date 2017-12-04 21:01
First, I would leave one core for other tasks like operating system functions.  I find this important in my case, because the machine becomes unresponsive to other tasks if I use all the cores.  I have an AMD machine with 64 cores.

Second, I also see a ramp-up in speed until the hash is fully in memory (I run on Ubuntu, but I guess the effects are the same).  As a work around for this, I usually analyze the first position in my list twice.

I guess you will do better with all engines if you turn HT off.
The latest AMD cores do well on HT, but I have not seem the same effect with Intel cores, according to what I have seen published.

If you look at the speed graph for Komodo on TCEC, you will see the same slow rise as you describe:
http://tcec.chessdom.com/live.php
Nobody knows why it behaves that way.
Up Topic The Rybka Lounge / Computer Chess / chess engine speed/settings on my new 48c workstation

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill