We learned a number of important lessons from our implementation of log skipping using range writes. First, we see that range writes are also useful for a file system journal. Under certain workloads, journaling can induce a large rotational cost; freedom to place transactions to a free spot in the journal can greatly improve performance.
Second, with read traffic present, the improvement seen by Bark is lessened but still quite noticeable. Thus, even with reads (in the uncached case, they comprise roughly one-third of the traffic to the main file system), flexible writes to the journal improve performance.
Finally, we should note that we chose to incorporate flexible writes underneath the file system in the simplest possible way, without changing the file system implementation at all. If range writes actually existed within the disk, the Bark layer would be much simpler: it would issue the range writes to disk instead of using a model to find the next fast location to write to. A different approach would be to modify the file system code and change the journaling layer to support range writes directly, something we plan to do in future work.