we export GPFS via NFSv3. The NFSv3 clients always request FILE_ASYNC NFSv3 writes. It looks like this slows down write performance - I assume that in this case GPFS' behaviour is same as with O_SYNC writes, i.e. the write returns when the data actually is on disk (?).
Is there any way to make the GPFS/NFS server to fake the FILE_ASYNC behaviour and acknowledge immediately once the data is in memory (and not yet on disk) ? I'm aware on the security implications, possible file corruptions, ...
Background: We run VMWare ESXi servers with NFS storage for the virtual machines' disks (using CTDB/GPFS). ESXi always requests FILE_SYNC NFSv3 writes, I can't change this on the client side.
