Performance

   
Figure: Read performance of the JNFS system vs. NFS on several different file sizes.

Performance in our simulations was reasonable for a Java application, but not nearly as good as NFS. On average, JNFS file transfers take about 8 times as long as analogous NFS transfers on a Sun SparcStation 10 running Solaris 2.5. We present a comparison of JNFS and NFS read speeds on various different file sizes in Figure 2. There are several factors that lead to worse performance in our simulation:

• The JNFS server runs as a user-level daemon over a native file system. Since it does not run as part of the kernel, it incurs a large expense in traps to the kernel. Similarly, the client also runs in user space, which further worsens performance.
• Since the Java byte-codes are interpreted instead of executed natively, actual execution of the file system code is much slower. On an NC, NFS would be slower because it would be implemented in Java instead of a compiled language.
• The RMI networking protocol has a large overhead in comparison with NFS's UDP-based implementation. Although many of the JNFS functions transfer a block of data across the network, the RMI service still marshals and un-marshals this data, resulting in more copying than necessary.