From: Miklos S. <mi...@sz...> - 2007-04-11 09:09:56
|
> > We've written a fuse-based filesystem that's used for storing millions > > of smallish files on a storage server. The fuse filesystem is then NFS > > mounted to dozens of machines. We're noticing that on occasion fuse will > > chew up all the CPU time on the server, processing hundreds of thousands > > of stat requests per second. > I've seen a similar issue with encfs on top of a NFS mount point (which > is sort of the opposite order to you). The server ends up crashing with > OOM kernel errors days to weeks later (I kid you not). i.e. I mount a > NFS server, run up encfs on that mount point, copy tonnes of data to it, > then stop using it. Then days to weeks later the server *always* > crashes/hangs via a OOM error. Umounting encfs and then NFS after > using it appears to stop the fault triggering. That is weird! Yes, NFS over FUSE does have problems, but I would have thought they were more benign, like getting "Stale file handle" errors once in a while. > Google shows a few people reporting issues between fuse and NFS in > the past? There was a bug on SMP, that would cause CPUs to go into an infinite loop one by one, until the whole machine completely froze. This was fixed in fuse-2.6.0, and actually has nothing to do with NFS other than the fact that the NFS server seemed to trigger this better than any other app. Miklos |