The Hardware Bottlenecks

An Experiment That Shows Bottlenecks

To demonstrate the various hardware and software bottlenecks that normally constrain networked file system write performance, we compared SunOS 4.1.3 NFS file write performance in various hardware and software configurations. We compare write throughput because it directly measures file system use of the network rather than internal caching. The results are summarized above.

All values are averages of 25 repetitions of writing a 64 megabyte (MB) file between two SPARCStation 20/71 workstations running SunOS 4.1.3 connected via a Myrinet.

Each bar on the graph and its associated bottleneck is described below.

Sync Ether

• NFS using synchronous disk writes which underuses both the network and disk bandwidth. NFS version 3 relaxes the synchronous write constraints. The bottleneck in this configuration is the use of synchronous disk writes at the server.

Async Ether

• NFS using asynchronous disk writes at the server on a 10 Mb/sec Ethernet. The network bottlenecks this system.

Async ATOMIC

• NFS using asynchronous disk writes and the ATOMIC LAN. Replacing the Ethernet with the ATOMIC LAN improves the performance to roughly 20 Mb/sec, which is equivalent to local file system performance. The disk is the bottleneck in this system.

No Disk ATOMIC

• NFS exchanging the write data only, without updating the disk.

Room to Improve

• No Disk Atomic throughput is significantly below the TCP process to process throughput, which implies that NFS protocol throughput can be improved. This is the base configuration for our improvements.

Go back to the ATOMIC-2 home page. / Go back to the ISI home page.

This page written and maintained by the ATOMIC-2 group.
Please mail us any problems with or comments about this page.
Last modified July 9, 1997