Re: [fuse-devel] fuse performence

[Miklos S. <mi...@sz...>]

On Thu, 1 Jan 2009, Kevin wrote:
> Hi, all
> 
> I have some observation on fuse performance I hope someone can explain to me.
> 
> One thing i observed is that fuse file system performance
> degenerates significantly over time if you run heavy load against it
> for long time. it could go down to 60-80% after a while.
> 
> Fuse user land code intrduced the "path lock" to replace the rwlock
> treelock in 2.7.x,  with this mechanism, if a parent path is locked,
> access to the sub tree will have to wait, why is this better than
> previous implementation? it seems that fuse spends lots of time on
> "wait_on_path", so if the root node is locked, access to any node
> would have to wait. it doesnt make much sense to me.

It is a reader-writer locking mechanism.  This means, that only
rename() and unlink() will actually block other operations.  Anything
else can run in parallel.

The old locking was the same only there was a single lock, so for
example a rename would prevent any other operation on the whole
filesystem.  The new locking is fine grained and only blocks
operations which actually depend on the path to be renamed/unlinked.

Thanks,
Miklos