[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
[ESPResSo-users] The myth of checkpointing
From: |
Axel Arnold |
Subject: |
[ESPResSo-users] The myth of checkpointing |
Date: |
Sun, 25 May 2014 20:28:21 +0200 |
User-agent: |
Mozilla/5.0 (Macintosh; Intel Mac OS X 10.9; rv:24.0) Gecko/20100101 Thunderbird/24.5.0 |
Hi all,
please note that checkpointing never really worked in ESPResSo, even
when using invalidate_system or similar. Actually I removed
invalidate_system, as neither the name was reasonable nor was it doing
anything useful. There are now two replacements:
- a command sort_particles, which should work for checkpointing, but
there are more caveats to that, see below.
- a flag recalc_forces to integrate which enforces force recalculation.
This you should use for timing purposes, when you want to recompute the
very same forces several times.
As for checkpointing, you should consider the following:
- the order of particles is important, as the random numbers are
distributed sequentially. The new sort_particles command really sorts
the particles not only across nodes as invalidate_system did, but
also on each node, which was missing.
- blockfile floats are not binarily exact, as they can be stored with at
most 17 digits precision. Therefore, you need to write binary
checkpoints. There is since ages a command for that, writemd, and now it
even has documentation. However, it can only store positions, velocities
and forces at present.
- other _integer_ information like type or bonds can be stored in a
blockfile safely, but interaction parameters, time step and all other
float information should be restored by the very same Tcl command in the
same order that you used to create them the first time. Only by that all
internal values will be binary equal.
- You don't have to write out forces. If you do and want to reuse them,
then you need to specify "reuse_forces" to the very first integrate
command after reading the checkpoint. If you don't write them, they will
be recalculated, which can lead to wrong results with LB.
- LB has its own checkpointing, but implies that you should write out
forces.
For more details, see the updated user's guide.
Best,
Axel
--
JP Dr. Axel Arnold
ICP, Universität Stuttgart
Allmandring 3
70569 Stuttgart, Germany
Email: address@hidden
Tel: +49 711 685 67609
[Prev in Thread] |
Current Thread |
[Next in Thread] |
- [ESPResSo-users] The myth of checkpointing,
Axel Arnold <=