Not logged inRybka Chess Community Forum
Up Topic Rybka Support & Discussion / Rybka Discussion / Rybka Cluster in Chessbase cloud
1 2 3 4 Previous Next  
Parent - By Barnard (Bronze) Date 2012-08-07 23:31
even if i dislike rybka and vas,i must admit that the cluster is really well hardware designed,and also the cluster software is really well implemented

and that it isnt or it can be related to any other engine

so in that point,i must tell 'chapeau' for both of them;seems that they work very well together,and that is really rare to fin
Parent - - By suj (***) Date 2012-08-13 14:49
hardware is something most people can buy but to optimise the hardware with the software is the real deal! Even doing this every cluster will have a optimum limit after which the performance will drop due to the overheads being greater.
Parent - - By Banned for Life (Gold) Date 2012-08-13 17:28
Even doing this every cluster will have a optimum limit after which the performance will drop due to the overheads being greater.

Sure, the software will be designed based on the number of nodes in the cluster, and a cluster with thousands of nodes will require different algorithms than one optimized for hundreds of nodes. And of course nobody will work on the thousand node problem until they have the hardware, which in the case of chess won't pay for itself...
Parent - By suj (***) Date 2012-08-14 10:30
The usual theory is more nps adds to strength but that is not the case with more nodes and mega mnps(1000+Mnps).

A smaller cluster actually outperformed the bigger cluster in my tests.For sure the software has to be designed keeping in mind hardware used. I have to say sjeng performs quite decently on commodity hardware(AKA desktops-GCP ran laptops using wifi in the past,Jens too uses desktops) and obviously performs strongers on servers.

Surely there is a strength gain from using infiniband or 10G in tournament play to the normal use of gigabit cards on playchess.
Parent - - By siah (***) Date 2012-08-08 19:46
You are always struggling about rybka's power and houdini's. shut up!
Parent - By Barnard (Bronze) Date 2012-08-08 19:47
i will shut up :smile:

all that a lady asks to me,is a good asking :wink:
Parent - - By suj (***) Date 2012-08-14 10:37
why johnny >>> seriously sjeng is not far behind.

Using sjeng cluster in centaur mode can win you freestyle events.

Also just using the rybka cluster in freestyle mode is not going to win you titles.Its the centaur skills and all the other engines he uses in combo with the cluster.

As a standalone engine sjeng is weaker than rybka/houdini but the scaling and hardware gives you the strength it needs. If GCP had more time at his disposal may be the engine strength can be further improved. Only time will tell...
Parent - - By Stonehenge (***) Date 2012-08-14 11:13

> As a standalone engine sjeng is weaker than rybka/houdini but the scaling and hardware gives you the strength it needs.


Software trumps hardware.

Houdini 1.03a on 1 core was about as strong as Deep Sjeng on 8 cores.
Houdini 3 on 1 core will be about as strong as Houdini 1.03a on 8 cores.

This means that Houdini 3 on 1 core will be about as strong as Deep Sjeng on 64 cores.
Could be an interesting match ;).
Parent - - By suj (***) Date 2012-08-14 13:03 Edited 2012-08-14 13:09
which sjeng are you talking about? You might be referring to sjeng wc2008 which is 4 years old or the sjeng ct2010 (2+year engine) which is a 32 bit single core engine which was a free engine that came with a magazine subcription.

>Software trumps hardware.


Totally agree on this.

>Houdini 3 on 1 core will be about as strong as Houdini 1.03a on 8 cores.


This is interesting!
Parent - - By Stonehenge (***) Date 2012-08-14 13:12

> which sjeng are you talking about?


Deep Sjeng c't 2010 32b as shown in the IPON list http://www.inwoba.de/bayeselo.html .
Parent - - By suj (***) Date 2012-08-14 13:13
ok thats too old and too weak:)
Parent - - By Stonehenge (***) Date 2012-08-14 13:28

> ok thats too old and too weak:)


True, and therefore a perfect illustration of the "software trumps hardware" line.
Parent - - By suj (***) Date 2012-08-14 13:36
It implies there are stronger private versions than the ct2010 within the sjeng team.
Parent - - By tomgdrums (****) Date 2012-08-14 13:41

> It implies there are stronger private versions than the ct2010 within the sjeng team.


Just out of curiosity, is Sjeng forever off the market?  Or will Gian Carlo think about selling it again? 

Given all the nastiness that can occur in the computer chess world it is understandable why he would keep it private.
Parent - By suj (***) Date 2012-08-14 14:01

>Given all the nastiness that can occur in the computer chess world it is understandable why he would keep it private.


You never know how the future will pen out but may be GCP will release it maybe not:cool:
Parent - - By Stonehenge (***) Date 2012-08-14 14:58

> It implies there are stronger private versions than the ct2010 within the sjeng team.


Who cares, private engine versions are vaporware.
The same can be said of Houdini 3 until it's actually released.
Parent - - By suj (***) Date 2012-08-15 11:36
GCP is happy to keep it that way and thats his choice.
Parent - By Uly (Gold) [mx] Date 2012-08-15 22:07
Yes, but it sucks, it's the story of Fruit or Zappa/Rondo, which will remain buried in some guy's basement because they ended on the hands of people that don't like to share.
Parent - By Barnard (Bronze) Date 2012-08-14 21:34
if you are a good player with good skills,of course,it is an advantage over the engine without a centaurr
Parent - By InspectorGadget (*****) [za] Date 2012-08-08 22:57
Alan, the problem here is that Barnard wants Rybka Cluster not to be what it is and want it to be something else. It will be the same wanting Hydra to be something else. Maybe wanting Hydra's software to run on equal hardware with Houdini. These pepole should rather compare Rybka 5 UCI when it is out with Houdini 3. Otherwise they just want the Rybka Cluster not to be what it is created for.
Parent - By tomgdrums (****) Date 2012-08-07 21:39

> The relevant data is the strength of the cluster on 64 or 200 cores not 1.


I know.  But that hasn't really been released either has it?
Parent - - By suj (***) Date 2012-08-13 14:43

>Personally, for me to even consider renting time with the cluster (even on the Fritz cloud, which you can get outbid which is weird!)


Not really you can get the reliable rate which is around 1.5 ducats per minute. Lets say you get 300 ducats for 30 euros and if its 1.5 ducats per minute that works out to 200 minutes for 30 euro.
1 minute is .15 Euro and this cloud has just made rybka cluster usable to all users. Imagine 64 cores say using 16 cores machine it would mean 4 nodes atleast and am sure Lukas would hardly gain anything after paying for bills (electricity).

With regards to all relevant data each person could have their own things to test and how is it going to be feasible for Lukas to guess what each one of us want and this theory always keeps changing. Best way is get 100 ducats have a list of things you want and get testing.

@werewolf I did see the 3 mnps and I had the same issue with houdini on engine cloud. I think its probably a buggy gui and once a reboot is done its fixed.

Kudos to Lukas for adding the cluster onto the chessbase cloud.
Parent - - By tomgdrums (****) Date 2012-08-13 14:54
I am sure Lukas is not really making money on the chessbase cloud option but at the same time after 9 hours of use (give or take a minute) I have paid more than I did for Rybka 4 UCI and I have to keep renting to use the cluster.

Again, I am sure the Cluster is the top chess "entity" (although a match would be cool and interesting!) I just can't justify the expense, YET, for my chess computing needs.  I need more autonomy than renting when available.
Parent - - By suj (***) Date 2012-08-13 15:03
thats how you see it but you are not comparing apples for apples.
Parent - - By tomgdrums (****) Date 2012-08-13 15:56

> thats how you see it but you are not comparing apples for apples.


You are correct.  That IS how I see it.

I just don't see the extra power of the cluster as being worth the money at this time.  Maybe that changes, maybe it doesn't.

Maybe if I was Topalov getting ready for a World Championship match I would need a super computer to help with preparation.  (  although he still lost.  :)  )
Parent - - By suj (***) Date 2012-08-14 10:19 Edited 2012-08-14 10:21
May be Topalov didnt really know how to use the cluster. If only I had topalov's chess knowledge and access to rybka and sjeng cluster's for preparation!

Also cluster then was probably really raw but perhaps it didnt give the rybka team enough time to work and polish on it as needed.All this is speculating anyway. Only people who were there will really know.
Parent - - By Kappatoo (*****) [de] Date 2012-08-14 10:54

> May be Topalov didnt really know how to use the cluster.


Are you joking? If no, what makes you think so?
Parent - - By suj (***) Date 2012-08-14 13:07
For sure he might not had time to operate the cluster himself. I am guessing he would have got his seconds and team to do the analysis not him doing things from scratch.

For sure Lukas or Vas would be more knowledgeable to use the cluster than topalov. So if Vas or Lucas had the chess knowledge as topalov and they are better users of the cluster they for sure would have performed much better.
Parent - - By Kappatoo (*****) [de] Date 2012-08-14 13:54
Of course I don't know any specifics about Topalov's work with the cluster. But I think both his and Anand's opening preparation fo the match were outstanding.
Parent - By tomgdrums (****) Date 2012-08-14 15:08

> Of course I don't know any specifics about Topalov's work with the cluster. But I think both his and Anand's opening preparation fo the match were outstanding.


I think the main point is that renting the cluster for opening preparation, while maybe giving some insights will not necessarily lead to better OTB chess performance.  (of course that can be said for any engine)
Parent - - By Gaмßito (****) [cr] Date 2012-08-07 08:59

> It is much stronger than Rybka 4.1.


If there is not too much to ask, please answer: how much stronger?

Gaмßito.
Parent - By Uly (Gold) [mx] Date 2012-08-07 11:23
And not only that, also, how was it tested? Lukas has a record of running tests for nodes reflecting Rybka real playing strength, and for hash adding 5 elo per doubling, in where after the conditions of the tests were known, it left a bad taste in my mouth (the former being done several years ago and for specific conditions, and the latter only revealing hash can have such an effect if one side has very low hash [2MB] and the other x4 but still very low [8MB], but nothing in-between, and nothing that would say that using more than 64MB would have an improvement).

The conditions need to be known also.
Parent - By Geomusic (*****) Date 2012-08-08 07:22
You mean like a Houdini Cluster? hehe  I'm sure Houdart has toyed with it. Maybe someday.
Parent - - By patrick delaurentis (**) Date 2012-08-08 22:24
couldent the cloud be used to gather mass high depth moves from all who use it and vas can share them with us vs keeping them for his own improvements?
Parent - By Uly (Gold) [mx] Date 2012-08-09 04:50
I think analysis made by people should remain the property of whoever rented the Cluster at the time and that Vas and Lukas don't have a right to look into it, it should remain private unless the person doing the analysis is happy with it being public.
Parent - By suj (***) Date 2012-08-13 15:00
thats what the chessbase feature something check does to see all possible moves in a set position.
Parent - By M ANSARI (*****) [kw] Date 2012-08-08 07:21
Yes of course.  If previous versions of Rybka are any indication it is that they perform best as the hardware racks up.  I remember seeing +100 ELO in results by just simply keeping everything the same (TC,test suites, opponents) and just moving the hardware from dual core to octa cores.  I would expect the same thing here, the bigger the hardware the better the performance against opponents.
Parent - - By Banned for Life (Gold) Date 2012-08-06 14:01
Vas and Lukas have spent the last few years optimizing the engine software and hardware to allow scaling up to hundreds of nodes and maybe more, and you want to throw this all away to test with 1 CPU? :lol:

If you want an engine optimized for one use one core, use Komodo. If you want an engine that is optimized to run well on up to 16 cores, use Houdini on a shared memory computer. If you want to use an engine optimized to scale to 64 cores or more on computers without shared memory, you'll want to use the Rybka Cluster. Running this on one core destroys the premise of the exercise and is a severe handicap and it indicates absolutely nothing as the cluster is not designed to run efficiently under these conditions.
Parent - - By Barnard (Bronze) Date 2012-08-06 21:54
Alan,Houdini is able to run on a 32 core machine,not only 16
Parent - - By Banned for Life (Gold) Date 2012-08-06 22:49
And if there were 64 or 128 or 256 core SMP machines, Houdini might run on those too. But at some point the shared memory model ceases to be a good way to go to split up the search, and a distributed method must be used.
Parent - - By Barnard (Bronze) Date 2012-08-06 22:53
im not an expert with hardware,i think you are right,but i cant speak with good knowledge

but i thinka that Houdart choose that model (shared memory) because almost all the people,havent a computer better than 32 cores,so he focused on the market that he can make money,and never tried the other model because at that model,hasnt market to make money (i dont know the number of people that will have computers with more than 32 cores,but i think is very ridiculous compared to peolpe with computers with 32 cores or less)
Parent - - By Banned for Life (Gold) Date 2012-08-06 22:57
An engine based on shared memory is both more market responsive, and also easier to develop. It's not surprising that this is the platform engines are developed for. But if you want to be able to scale the cores into the hundreds, a different parallelization strategy may be necessary...
Parent - By Barnard (Bronze) Date 2012-08-06 23:39
well,i dont know any persoon soo fool to do it,i mean build a cluster just to play blitz games at playchess,or have an almost sure win in the cc tournaments

but i must admit that if i were billionaire,i will buil the same supercomputer that Topalov used,and i will contract Houdart in full tiem to develop a version of Houdini able to play in 8192 core cluster...just to defeat rybka cluster :twisted:
Parent - - By Geomusic (*****) Date 2012-08-08 07:31
what about making a cpu by adding or subtracting energy states from electrons using it like a quantum abacus?
Parent - - By Labyrinth (*****) [us] Date 2012-08-08 11:29
How exactly would that compute?

This kind of thing already is used in photodetectors and lasers.
Parent - By Geomusic (*****) Date 2012-08-09 05:34
by using a form of boolean algebra which uses an electron shell base instead of a base 10 or binary base. Not sure exactly how it might work :)
Parent - By suj (***) Date 2012-08-14 10:24
Almost all engines above the 40 core mark will struggle.Huge drop off almost like falling off the cliff.

Zappa was probably the only engine but to be honest the time control I ran with for KOE was 180/0 and that time zappa was the strongest engine in long time control followed very closely by rybka (rybka 1.0) days.
Parent - By tomgdrums (****) Date 2012-08-05 21:43

> As for the testsuite...the Clusters results were impressive, but I couldn't finish testing and it was failing on a few positions at the end. Someone outbid me and was willing to pay over 10 ducats a minute :eek:


The bidding thing is why I didn't buy DF 13.  I don't want to worry about bids.  It is why I never buy anything on ebay.

I would have performed horribly in the age of bartering.  :)

Although i did cave and purchase Fritz 13 (my first, and only my second, Fritz GUI since Fritz 10 and I am pleasantly surprised how easy it is to analyze and annotate with it.)  And chessbase may have even outdone the ShredderClassic GUI with regards to engine management!!
Parent - - By Cole [us] Date 2012-10-22 22:04
I hope you still plan on running this H3 vs cluster match and wish you well in doing so!  Any chance you would post the games?
Parent - - By Carl Bicknell (*****) [gb] Date 2012-10-23 05:23
There's no news at the moment. But ANYONE could run this match...Houdini 3 is public ally available and so is the cluster.
Up Topic Rybka Support & Discussion / Rybka Discussion / Rybka Cluster in Chessbase cloud
1 2 3 4 Previous Next  

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill