Micro-benchmarks

  We use the Bonnie benchmark [1] to evaluate performance when writing and reading a large file. To eliminate caching effects we use a 512MB file, twice the size of main memory. In our results we also include the performance of FFS and NFSv3 for completeness.

  

Figure 5: Bonnie Sequential Output (Block). DisCFS throughput is comparable to NFSv2 throughput.

  

Figure 6: Bonnie Sequential Input (Block). DisCFS throughput is less than 50% of NFSv2 throughput, mostly due to its lack of prefetching.

  

Figure 7: Bonnie Sequential Output (Rewrite). DisCFS throughput is closer to NFSv2 throughput than for reads, but still suffers from its lack of prefetching.

Figure 5 presents results for block writes. The perfomance of DisCFS is equivalent to that of NFSv2 because the credential related overheads of DisCFS are amortized over the time of the experiment. Remember that this experiment is on a single file.

Figure 6 presents results for block reads. We observe that for read operations DisCFS trails NFSv2 by more than 100%. The reason for this behavior is that NFSv2 prefetches blocks, an optimization which we have yet to incorporate in the current implementation of DisCFS.

Finally, in Figure 7 we experiment with the cost of rewriting blocks. More specifically, in this test Bonnie reads a block, dirties it, and then writes it back. As expected, the performance of DisCFS is closer to NFSv2 than in the read benchmarks, however DisCFS still suffers from its lack of prefetching.