direct vs. in core vs. disk i/o computations

Jochen Küpper jochen at unc.edu
Fri Sep 6 20:39:11 BST 2002


On Fri, 6 Sep 2002 08:58:24 +0200 Dave Moore wrote:

Dave> "Since MOLPRO does not seem to allow you to keep integrals in
Dave> memory, even when there is plenty of space, you have to trick
Dave> MOLPRO into writing the intergral table into memory by using a
Dave> ramdisk. Just in case you are unfamiliar, a ramdisk is an area
Dave> of memory that is mapped to have file access like a disk drive
Dave> ... sort of the inverse of virtual memory."

But the ramdisk is going through multiple filesystem layers, for
example on linux these could be: ext2 (or reiserfs, or ext3), VFS for
the write, then both again for the read.

Dave> I have tried this approach to test a couple of simple examples
[...]
Dave> Perhaps someone with more experience could chime in here?

It should work up to the limit for the ramdisk. VFS should not be a
problem here, but I am not sure the 4GB adress space of IA32 can be
overcome in the current ramdisk implementation.

On Fri, 06 Sep 2002 13:33:08 +0100 Peter Knowles wrote:

Peter> The philosophy has been that on modern computer systems, it
Peter> would be largely irrelevant, since many operating systems
Peter> efficiently cache the file system in memory, if that memory is
Peter> available.

This is also using the expertise of kernel developers and should be at
least as fast as mmap'ing a file, whch is in turn at least as fast as
the ramdisk. So theoretically it is at least as fast as the ramdisk.
In practice I guess the OS's are better than MOLPRO could become with
any reasonable amount of work. I'd say Peter & Co better spend their
time on analytical gradient for RCCSD(T) (or so) <wink>.

One advantage of the ramdisk is that the mem will be in physical
memory as much as possible, if you make the ramdisk non-swapable. The
caching will be trashed by any job that needs memory. In practice you
are concerned about dedicated machines, though, so that is a non-issue.

Greetings,
Jochen
-- 
University of North Carolina                       phone: +1-919-962-4403
Department of Chemistry                            phone: +1-919-962-1579
Venable Hall CB#3290 (Kenan C148)                    fax: +1-919-843-6041
Chapel Hill, NC 27599, USA                            GnuPG key: 44BCCD8E




More information about the Molpro-user mailing list