Dynamic Availability

Finally, by simulating dynamic availability, we examine how often users or applications will be oblivious that D-GRAID is operating in degraded mode. Specifically, we run a portion of the HP trace through a simulator with some number of failed disks, and record what percent of processes observed no I/O failure during the run. Through this experiment, we find that namespace replication is not enough; certain files, that are needed by most processes, must be replicated as well.

In this experiment, we set the degree of namespace replication to 32 (full replication), and vary the level of replication of the contents of popular directories, i.e., /usr/bin, /bin, /lib and a few others. Figure 3 shows that without replicating the contents of those directories, the percent of processes that run without ill-effect is lower than expected from our results in Figure 2. However, when those few directories are replicated, the percentage of processes that run to completion under disk failure is much better than expected. The reason for this is clear: a substantial number of processes (e.g., who, ps, etc.) only require that their executable and a few other libraries are available to run correctly. With popular directory replication, excellent availability under failure is possible. Fortunately, almost all of the popular files are in ``read only'' directories; thus, wide-scale replication will not raise write performance or consistency issues. Also, the space overhead due to popular directory replication is minimal for a reasonably sized file system; for this trace, such directories account for about 143 MB, less than 0.1% of the total file system size.