I like your goal, as long as you standardize on a cluster (perhaps 64 cores), since most of the strong programs support clusters and there is no point on crippling them. 5 of the 8 programs in the ICGA 19x19 contest today are using clusters. I'm not, because my cluster code is weaker than 12 way SMP today. One is using 56 core SMP, so it doesn't need a cluster. And one program is running on a single core.
David From: [email protected] [mailto:[email protected]] On Behalf Of Adrian Petrescu Sent: Wednesday, September 29, 2010 5:56 PM To: [email protected] Subject: Re: [Computer-go] Homogenous environment for Computer Go tournaments Hi David, That's a lot to think about, I'll come back to you about it soon. But I should mention that I was not at all considering having everyone "use whatever Amazon resources they could pay for," which of course leads to the same type of arms race as before. My idea was always to get the community to agree on a baseline that's both powerful enough to be interesting and cheap enough to be affordable (possibly the 64 cores you mentioned) and just set that as the entrance fee (minus whatever subsidy the organizers can get). I'll have a fuller response after I examine the Windows offerings more. Thanks! -Adrian On Wed, Sep 29, 2010 at 5:39 PM, David Fotland <[email protected]> wrote: They do have some HPC support: http://aws.amazon.com/ec2/hpc-applications/ But currently it is limited to 64 cores, and linux only. They don't have pricing on the web that I could find. However, if you allow everyone to use whatever Amazon resources they can pay for, doesn't that defeat the purpose of using standard hardware? Some programs will run on the HPC 64 core, and some will run 8-way SMP. As long as the Amazon HPC is available, then it seems that large clusters are available to any programmer, so there is no need to force everyone to run on Amazon. Mogo can use its 160 core cluster, and any program can use an Amazon 160 core cluster. I don't think it would fair if one programs says "I didn't have time to develop MPI code so I can't run on a cluster" and because of that, prevent everyone else from running on a cluster. If you want to standardize on equivalent hardware for everyone, I think it should be a 64-core cluster (say 8 by 8-cores). Then it becomes a programming contest again. But don't limit the contest to SMP just because some programmers don't want to develop cluster code. Cluster code that scales well is very difficult. David From: [email protected] [mailto:[email protected]] On Behalf Of Adrian Petrescu Sent: Wednesday, September 29, 2010 4:56 PM To: [email protected] Subject: Re: [Computer-go] Homogenous environment for Computer Go tournaments Hi guys, glad to see all the feedback :) Petr, AWS provides a virtual machine, but the software image running on it can be endlessly customized. I'm very confident Pachi can be built for it with minimal headaches. Oliver, the per-hour cost for the "High-CPU Instances" that I would consider the most reasonable for such an event cost $0.68/hour for Linux images and $1.16/hour for Windows images. Machines can be spun up for single hours at a time, there's no "minimum reservation", so the entire event could be run at about the cost of $20 per person. And that's aside from the fact that I feel we have a chance to get a discount or donation from AWS if we go with them. David, Windows Server 2008 is one of the default supported instance types for AWS (details here: http://aws.amazon.com/windows/). The prices are slightly higher (about $0.50 more per hour) to offset the cost of the Windows licenses, but it's still very much within reason. So you can absolutely run Many Faces without the need to put your source code onto the cloud. That is, of course, one of the first things I considered :) Cheers, Adrian On Wed, Sep 29, 2010 at 4:44 PM, David Fotland <[email protected]> wrote: For a big SMP or cluster, the program is tuned somewhat for the environment. If the OS is unusual it might be necessary to compile the program specifically for that machine. Both of these are large barriers to entry. I won't put my source code into the cloud to be compiled there. My engine only runs on Windows or windows emulation since I use windows threading, not posix threading, and I use the Microsoft MPI library. If the cloud system you are using is running Windows server 2008, then I could participate. Otherwise your proposal to standardize hardware may just become a way to exclude commercial programs. Regards, David From: [email protected] [mailto:[email protected]] On Behalf Of Oliver Lewis Sent: Wednesday, September 29, 2010 4:47 AM To: [email protected] Subject: Re: [Computer-go] Homogenous environment for Computer Go tournaments Adrian Can you give us a rough estimate of how much it would cost, per entrant, to run the programmes on a heavyweight hardware configuration (of your choice)? Oliver On Wed, Sep 29, 2010 at 12:52 PM, Petr Baudis <[email protected]> wrote: Hi! On Tue, Sep 28, 2010 at 12:04:10AM -0700, Adrian Petrescu wrote: > I'm not much of a participant in the field of computer go, but I am an avid > observer, so it puzzles me when I see things like the recent 9x9 "World > Championships" being plagued by issues of operator error, hardware > malfunction, network outages, etc. Even when everything goes smoothly, it's > hard to take the results too seriously when some programs are running on a > 16-core dedicated machines, and others are running on the developer's > personal laptop. I believe "hardware-open" tournaments are great and should not go away. But I also think "hardware-fixed" tournaments certainly do have their place and could provide very useful feedback. These tournamens do happen sometimes, e.g. the computer go tournament in Tampere this summer was with fixed hardware. (Well, almost. But at least no clusters.) If someone would organize a tournament e.g. using the EC2 hardware platform (and if the OS is UNIXy), Pachi would participate. -- Petr "Pasky" Baudis The true meaning of life is to plant a tree under whose shade you will never sit. _______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go _______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go _______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
_______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
