On Mon, Dec 6, 2010 at 9:39 AM, Chris Garrison <[email protected]> wrote: > AFS gurus, > > We have a professor who's hoping to implement our AFS cell (which as a > service, we call RFS, the Research File System) as his department's solution > for user home directories. After working on this awhile, he thinks there > may be a show-stopper. Is there an problem with using AFS with applications > that use mmap to solve IO-bound issues? > > The problem is a esoteric issue called memory mapping. We have > applications that we use extensively in geophysics that use memory > mapping to improve performance. The relevant C function is "mmap" (run "man > mmap" on any unix or macos machine). Memory mapping is used to reduce seek > times when a program does a lot random access IO, and is a well known > performance trick for IO-bound applications. > > Now I haven't written a simple test program to isolate the problem to > this specific issue, BUT I know that applications I use daily and that I > know use memory mapping fail when used on data stored on rfs. The same > application and data work fine when the files are local disk files or > served by nfs. Thus the memory mapping hypothesis is a strong possibility. > > Is this a know limitation? If so, is there a solution? > > We are running OpenAFS 1.4.11 on the server, and I am not sure what he is > running on the client side, but I'd imagine it's the latest stable one on > the openafs.org site. I'm also not sure which type/version/kernel of Linux > he's running on the server end.
Asking would be better than imagining as some versions of the Linux client had issues with mmap() > > Thank you for any help/suggestions. > > Chris > -- Derrick _______________________________________________ OpenAFS-info mailing list [email protected] https://lists.openafs.org/mailman/listinfo/openafs-info
