> -rw-r--r-- 1 dave dave 252901 2012-04-10 08:33 New/1.rst
> -rw-r--r-- 1 dave dave 252901 2012-04-10 08:33 New/2.rst
> -rw-r--r-- 1 dave dave 252901 2012-04-10 08:33 New/3.rst
> -rw-r--r-- 1 dave dave 245309 2012-04-10 12:59 New/4.rst
This is the problem - 4.rst is shorter than the rest of the restart files.
This is a problem that has only appeared recently - this code has worked
fine for many years. Something has changed in the way certain versions of
Linux uses write buffers - It is extremely dangerous and I am bewildered why
specific versions of Linux would choose to do this - It seems the buffer is
only ever flushed if the file is closed.
I would like to start documenting specifically which versions of Linux do
this. Also does the same thing happen with mdcrd and output files (which
also aren't closed during a run). It seems to me there are dangerous write
buffer settings being set by default.
There are two options to fix this:
1) Have pmemd (and probably sander too) open and close all files on every
write. This is undesirable since it will hurt performance a lot.
2) Find out what setting is being, rather foolishly set, on specific new
linux configs and then figuring out what /proc setting needs to be updated
to disable this. Then document that this needs to be set startup / login
scripts that have permission to edit /proc.
We can fix the restart issue with an open and close since one typically only
writes this every hundred thousand steps or so but I suspect a similar issue
is occurring with the mdcrd and mdout files that mean these might not match
where the restart file is in some cases. So it is not a very satisfying fix.
Can everybody who has seen this issue please post the specific version of
Linux they are running and we can try to see if there is a correlation.
All the best
Ross
/\
\/
|\oss Walker
---------------------------------------------------------
| Assistant Research Professor |
| San Diego Supercomputer Center |
| Adjunct Assistant Professor |
| Dept. of Chemistry and Biochemistry |
| University of California San Diego |
| NVIDIA Fellow |
|
http://www.rosswalker.co.uk |
http://www.wmd-lab.org/ |
| Tel: +1 858 822 0854 | EMail:- ross.rosswalker.co.uk |
---------------------------------------------------------
Note: Electronic Mail is not secure, has no guarantee of delivery, may not
be read every day, and should not be used for urgent or sensitive issues.
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Tue Apr 10 2012 - 12:00:05 PDT
