[CHiPPS] Questions on Blis

[Yan Xu]

As to gap, you are right. The best bound seems do not move. The parallel log was changed to style of commercial solvers last year, I did not test much. I will take a look.

For a child node, Blis only store the difference with its parent. I haven't thought about zipping and don't know how to do this yet. Any suggestion?

Thanks,
Yan


-----Original Message-----
From: Alessandro Tiberi [mailto:alessandro.tiberi at gmail.com]
Sent: Thursday, March 12, 2009 10:58 AM
To: Yan Xu
Cc: chipps at list.coin-or.org
Subject: Re: [CHiPPS] Questions on Blis

Yan, thank you for your answers.
> 1) The 0% printed in log is truncated. Its actual value is something like 0.0001%.
>  A separated issue is that the gap parameters of Blis doesn't work. Need fix.
>
I guess it should go something like this: when the master process
receives a solution whose value is between the gap, it  should send a
termination message to all the other processes, right?

> 2) For parallel runs, the best bound will be updated periodically, so sometimes it doesn't move.
>
This does make sense. However I could not find an example in which the
best bound is ever updated... do you happen to have any?
> 3) I tried to incorporate COIN's presolve, but couldn't make it work, so I quit. If you can contribute yours, that's great!
>
I am attaching a tgz with the three files I had to modify: BlisModel.h,
BlisModel.cpp and BlisTreeNode.cpp. My changes/addition are enclosed in
comments like // AT , so should be easy to spot.
I tested it only with clp as solver, but it seems to be fine (for both
serial and parallel build). If you want to test it remember
-Blis_presolve 1 :-)
> 4) Blis stores nodes by differencing to save memory, subtrees are sorted of compressed. Probably, it need functionality to store subtrees in hard drive.
>
Do you thing that adding a "compressing" wrapper  to the node
description would help? I was thinking of something like zipping the
description of not used nodes before storing them... if this makes any
sense. Thank you again!

Best,
Alessandro