Observations and Queries

From: Nicholas Robinson <nprREMOVE@THISbottlehall.co.uk>
Date: Sat Jun 25 2005 - 06:06:22 CEST

Hi Tomaz

I've finalised the configuration file for my real timetable - at the moment,
I'm using this year's allocations of teachers but should get the plot for
next years within a couple of days.

I tend to run tablix2 on a pvm cluster of two (old) machines using -n 10. I
think this may not be enough horse-power. I can add a third machine but it's
even older than the other two.

From my earlier email of 7/6/5, I've done as described in that email as
follows:

1. I've split the timetable events into two groups - one (called 'hard') which
includes all double lessons, the part-time teacher's events and all core
lessons where the two classes for each year have to have some subject lessons
at the same time. The other group contains the rest and I label this one
'easy'.

2. Running tablix2 on the 'hard' group results in a solution in just under two
minutes with fitness=0. This is easy to check and is 100% correct - and
generates very good quality timetables. In fact, it enables us to have an
additional set of double lessons that have never been possible before, for
just an extra 15 seconds execution time!

3. Editing a result xml file from the 'hard' run and changing 'time' to
'fixed-period' and including this with the 'easy' events becomes the
configuration file for the second run.

4. The second run, now I've got everything in, doesn't find a solution where
fitness=0 in a reasonable length of time - although it gets down to around
eight conflicts within a few minutes and eventually down to three after 30-40
hours. I'm not too worried because next year's plot will include two more
teachers than this year and overall should make a solution much easier to
find. No matter how many times I run it, the conflicts always arise in in the
same places, so it may still be an error in my configuration - or at least
how I edit the two bits together for the second run.

5. If I run the configuration file in one go without splitting the events up,
then it converges to a fitness where the number of conflicts is higher than
running it in two stages. Again, the conflicts mostly arise in the same class
as 4. but with the double lessons for this class.

So, it appears that, with the limited computing resources I have, running it
in two stages is better at the moment.

Does anyone have a feel for what would be a workable configuration of machines
for a real run? I could take over, say, twenty machines running Windows
(still haven't persuaded my headmaster to move over to Linux) and use a
morphix approach. Is it easy to build a morphix configuration for 0.2.1?
Is it feasible to run pvm over the Internet rather than just locally?

Thanks and best wishes

Nick
Received on Sat Jun 25 06:03:13 2005

This archive was generated by hypermail 2.1.8 : Tue Aug 16 2005 - 20:44:08 CEST