Jim, I can provide some answers to the fuse question. In principle, the fuse module in PVFS works the same way as the kernel module - i.e. there is still a user to kernel back to userspace transition. The tests that we did (about 4 years ago) showed that pvfs2fuse had similar performance to the kernel module as long as the block size was less than 128K.
John. On Mar 12, 2012, at 3:44 PM, Jim Kusznir wrote: > Hi all: > > I'm still on PVFS 2.8.2 on my cluster, but enough things have been > going wrong that I'm finally getting some traction toward upgrading. > Furthermore, we're beginning work on an interoperability project that > will make our pvfs2 storage available on another cluster. As I look > into OrangeFS, I'm finding that I'm full of questions, and so far > haven't found much useful documentation. There's the "high level > features", the list of open projects, but I haven't found much with > what currently works and how to build for various platforms. > > For example, on my cluster, all packages must be rpms. I've got a > spec file that I used to build both the pvfs2 userspace and a seperate > one for the kernel modules, but for pvfs-2.8.2 (It was actually for > several versions earlier that I bumped up a few times). At this > point, I suspect I need a new .spec file, and I'm not very good at > writing them. As I recall, there used to be a .spec file in the pvfs2 > source tarball, but I couldn't find one this time. Is there a .spec > file for building OrangeFS into rpm(s)? > > Second question: fuse support > It seems that the vast majority of my difficulty has come with the > kernel module. Furthermore, the cluster I'm trying to intertie with > has stated that they will NOT load a kernel module on that cluster. > My users of course have no clue about MPI-IO, and have no desire to > rewrite portions of their code to make use of it; they do all their > I/O via standard filesystem calls. So, is the kernel module still my > best/only option, or is there a fuse or other module? Is the fuse > module "faster/better" than the kernel module (by staying in > userspace)? > > Of cousre, the other cluster owner would prefer that I just provide an > NFS export for him to add to the automount table...Is that fessable? > > Presently my cluster headnode is a pvfs2-client, and it > crashes/reboots frequently as a result of I/O intensive activities on > it (eg, sftps, tarball creation/expansion on a pvfs2 volume, etc). So > re-exporting my pvfs2 volume as NFS from this node seems like a bad > idea (especially as it also hosts the user home directories for the > cluster). Also, in general it seems that my pvfs2 performance has > been pretty poor in general, especially from this node, and as such > I'm quite concerned about that. > > Finally, anything special I should be aware of with OrangeFS upgrade > and interoperability? > > --Jim > _______________________________________________ > Pvfs2-users mailing list > [email protected] > http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users _______________________________________________ Pvfs2-users mailing list [email protected] http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users
