I have noticed that I am getting about half the write throughput from ESXi to Solaris ZFS volume (20MB/s) over NFS then I do to a Linux EXT3 volume over NFS (40MB/s).
When writing to the ZFS volume from a Linux client with an async mount I get 110MB/s versus the Linux NFS server's 85MB/s, so async operations fly to the Solaris box. The backend storage is a hardware RAID6 (12 SAS disk) with 512MB of NVRAM cache and ESX issues all it's writes FSYNC to assure VM storage consistency. I'm NOT using the dangerous 'async' NFS server option on the Linux side, but I have the server option 'no_wdelay' set which prevents the NFS server on Linux from coalescing writes. Without this option writes to the Linux server are about the same as to the Solaris box. I figure the write delay on these synchronous writes has a halving affect on the throughput. This is a handy feature if the NFS client does asynchronous writes I suppose, but not so much for synchronous writes and especially not when going to NVRAM. What I am asking is does the Solaris NFS server coalesce writes? And if so is there an option to disable that? Per-export? Globally? -Ross _______________________________________________ Solaris-Users mailing list [email protected] http://www.filibeto.org/mailman/listinfo/solaris-users
