From the start we encountered problems with XLink's implementation. We were unable to use our test suite and could complete our testing only after we turned off both file cache and file locking within the Omni-NFS gateway.

Without file cache, Omni-NFS' performance numbers were extremely low: barely 1.5 MB per second. Omni-NFS' caching strategy treats every SMB connection independently. And each SMB connection's file table is independent, so changes made to one file table (such as creating a file) are not immediately seen by other SMB connections. Files created within one SMB share are available on other SMB shares only after Omni-NFS synchronizes those share's file table with the Unix server.

We also encountered problems with file locking. When enabled, we unearthed many cases where files locked using the NT file locking mechanisms would not unlock. Our testing with Coffee Computing's FileMetric 1.0 test suite rejected Omni-NFS' file-locking mechanism. When confronted with our findings, XLink conceded that the version we tested had known file-locking problems. It states these issues will be fixed in its next version.

More Problems Our problems didn't end with malfunctioning file locking and file caching. Omni-NFS' NIS support is poorly implemented. When we attempted to import our 5,100-entry password file, we found that only a fraction was imported. Although XLink promptly supplied us with fixed binaries, they aren't part of the standard distribution.

Once we were able to create files within Omni-NFS, we found attributes were set incorrectly (global read, write, and execute). This lack of attention to detail makes Omni-NFS a poor choice for an enterprise level NFS gateway.

Jeff Ballard can be reached at jballard@nwc.com.

How We Tested NFS Gateways
We ran each of these gateway products on an enterprise-class Windows NT server: a 200-MHz quad processor Pentium Pro Compaq Computer Corp. ProLiant 6500 server. We also used a Sun Microsystems Ultra 60, running SunSoft Solaris 2.6 with NFS version 3 as our back-end NFS server. On the client side, we used a collection of single- and quad-processor Pentium Pro workstations.

The network throughput required for our testing was extreme. Our network consisted of ATM edge switches with a 310-Mbps backbone (two 155-Mbps OC-3 connections). We connected both servers and six of our clients directly to the edge switches. The remaining machines were attached to an edge switch over a switched 100-Mbps Ethernet network.

Coffee Computing Corp.'s FileMetric 1.0 benchmark (www.coffeecomputing.com) was used to gather our throughput readings. The throughput testing stressed file reads through each NFS gateway.