From: Guillaume R. <Gui...@em...> - 2004-12-10 16:25:55
|
Dear Mark, Your mail is very interesting. We have just started to test Choco performance, since this was not our first requirement. But Sure, it is obviously a critical aspect of the solver. We made some improvement since the last release on the web site. This in on the CVS, but we will send you a jar directly (and publish it as soon as possible, next week we think) compared to the claire version of choco. If we try the NQueens problem on a 3 GHz machine (JDK 1.4.2_03), with the server mode and 13 queens, we find all solutions in 60 seconds approximatively, which is much more acceptable. Moreover we plan some improvements like observer lists for each variables, lazy iterator building, simplified propagation queue... but we still need to test these improvements to check their correctness. Another point is that the JDK 1.5 is quite more efficient with Choco (at least in client mode). So it may be a good idea to use this new version. Could just try again your benchmark with the new build we will send you separately ? Thank you for helping ! > 1. How much of this slowdown is because the Java version is relatively new > and not yet optimized, and how much is due to Java versus C++ speed? I think a lot of the slowdown was due to the first versions of Choco. The current CVS version is much more faster and we hope this will improve in future version thanks to the users help :). But there will be always an overhead due to the Java technology (even if SUN try to improve it as much as possible). > 3. Is Choco.sf.net likely to be a good option for us, if we want to use > just its integer solver, but extend it with hooks to report changes > in domains etc? Sure it is ! * First Choco was designed to be easily extendable and we experience it very often (to add new feature like new domain, new search (conflict directed backjumping last weeks), Oadymppac traces thanks to AspectJ...) * Moreover, we chose BSD license to allow everybody to use the project even in a commercial context. So you are welcome to use it (and certainly to help us to improve it !). Thank you for your mail, Best regards, Hadrien Cambazard and Guillaume Rochart |