[PVFS2-users] mmap() support?
rross at mcs.anl.gov
Fri Oct 21 17:53:46 EDT 2005
We definitely do *not* support shared mmap(). I would expect that we
will never ever support that -- it very much goes against the PVFS2
Regarding the delays, I don't recall any previous reports of this sort
of behavior. I agree that it is confusing. "less" is a read
application too, right? So there must be something different about
*how* vi and less are accessing the file.
Perhaps we should strace them and see what they're doing that is different?
Thanks for keeping us on our toes!
> We just upgraded to ver. 1.3.0, and, the problems we have had earlier
> are gone. Thanks as always for your hard work developing the package.
> However, we are seeing some new issues....
> First of all, one of our users reports problems when trying to use
> mmap() to operate on a file. I am wondering if this is something PVFS2
> does not support. In any case, attached is a program that is
> demonstrating the problem.
> To run, do:
> a.out <from_file_name> <to_file_name>
> the error we're seeing is reported as:
> mmap error for input
> mmap error for output
> Segmentation fault
> There are also two other issues.
> One was already reported - it's the messed up time stamps upon file
> The other is somewhat more subtle - we are seeing considerable delays
> (~5 minutes) when accessing the files on PVFS2, even small ones, with
> commands like vi, less,... Interestingly, using cat, more,... is pretty
> fast, which makes me to believe that the slowdown happens when the file
> is opened for writing. However, I haven't seen any delays on MPI-IO or
> UNIX I/O issued from a program, so, it's confusing. Does anyone else see
> this kind of delay?
> We're running (I believe) 12 I/O servers on x86-64 (SuSE 9.0, I believe
> - Brian, correct me if I am wrong). The clients are all x86-64 running
> SuSE 9.0 with 2.6.11 kernel.
More information about the PVFS2-users