NFS async writes - HEADSUP
Heads up folks. DragonFly NFS clients in master can now do async
writes and when the kernel flushes we're talking 64 16K packets at
least. On a fast GigE ethernet this can blow out the NFS server's
socket buffer (if the server isn't running the latest DragonFly master).
If such a situation arises for you, either use a TCP NFS mount instead
of UDP (-T option to mount_nfs, or "tcp" option in /etc/fstab), or
reduce the maxasyncbio's with a newly available sysctl called
vfs.nfs.maxasyncbio. The default is 64, which requires about 2MBytes
of sockbuf space on the NFS server.
Again, look forward to a bit of instability with NFS on the master
branch for the next few days as the kinks get worked out.
-Matt
Matthew Dillon
<dillon@backplane.com>