In fuse work I've done and seen, it is possible to extract decent
performance from fuse, but it isn't free. By default, FUSE only
performs I/O 4K at a time. With the direct_io mount option, you get the
benefit of only a single copy from userspace to userspace effectively
for I/O buffers and the buffers are then sized all the way up to 128K.
For my testing, the 128K buffers have been sufficiently large to extract
some decent throughput with. The 4K buffers with the additional
user->kernel copy (i.e. mounting without direct_io) have been more
problematic for my performance tests -- but that may just be the kind of
stuff I'm using fuse to accomplish. If you guys actually make BMI calls
directly in FUSE then I suspect you could get performance fine. I'm not
sure a simple wrapper of the system interface will get you there
completely. Not the kind of answer that you'd find super exciting --
but I still find FUSE to be very neat crap. If the system interface is
now more modular, as may be the case, you could utilize the state
machines and build a great POSIX interface I think.
The drawback of the direct_io flag is that you can't execute files in
any fuse volume mounted with direct_io.
Cheers,
Brad
On 02/01/2010 07:18 PM, John Chandy wrote:
Walt,
Its been about 3 years since I wrote the PVFS code and I havent done
any extensive testing to qualitatively pronounce it production worthy.
Sumit Narayan did some further work when he was at ANL during the 2008
summer and wrote a report about his testing at
http://www.engr.uconn.edu/~sun03001/docs/pvfs2fuse-rpt.pdf
. The basic summary is that FUSE works fine for small accesses but
does worse than the PVFS kernel implementation for large accesses (>
128K).
John.
On Feb 1, 2010, at 5:29 PM, Walter Ligon wrote:
So who knows the status of the FUSE interface? Is there any reason to
believe or not believe it is production worthy?
Walt
_______________________________________________
Pvfs2-developers mailing list
[email protected]
http://www.beowulf-underground.org/mailman/listinfo/pvfs2-developers
_______________________________________________
Pvfs2-developers mailing list
[email protected]
http://www.beowulf-underground.org/mailman/listinfo/pvfs2-developers
_______________________________________________
Pvfs2-developers mailing list
[email protected]
http://www.beowulf-underground.org/mailman/listinfo/pvfs2-developers
