We now measure the performance of unmodified Linux ext3 running on top of Bark. For this set of experiments, we mount the ext3 journal on Bark and let all other checkpoint traffic go to disk directly.
For a workload, we wished to find an application that stressed journal write performance. Thus, we chose to run an implementation of the classic transactional benchmark TPC-B. TPC-B performs a series of debits and credits to a simple set of database tables. Because TPC-B forces data to disk frequently, it induces a great deal of synchronous I/O traffic to the ext3 journal.
Table 3 plots the performance of TPC-B on Linux ext3 in three separate scenarios. In the first, the unmodified traditional journaling approach is used; in the second, Bark is used underneath the journal; in the third, we implement a fast ``null'' journal which simply returns success when given a write without doing any work. This last option serves as an upper-bound on performance improvements realized through more efficient journaling.
Note also that each row varies whether table reads go to disk (uncached) or are found in memory (cached). In the cached runs, most table reads hit in memory (and thus disk traffic is dominated by writes). By measuring performance in the uncached scenario, we can determine the utility of our approach in scenarios where there are reads present in the workload; the cached workload stresses write performance and thus presents a best-case for Bark under TPC-B.
From the graph, we can see that Bark greatly improves the overall runtime of TPC-B; Bark achieves a 20% speedup in the uncached case and over 61% in the cached run. Both of these approach the optimal time as measured by the ``null'' case. Thus, beyond the simulation results presented in previous sections, Bark shows that range writes can work well in the real world as well.
Figure 13 sheds some light on this improvement in performance. Therein we plot the cumulative distribution of journal-write times across all requests during the cached run. When using Bark, most journal writes complete quickly, as they have been rotationally well placed through our simple skipping mechanism. In contrast, writes to the journal without Bark take much longer on average, and are spread across the rotational spectrum of the disk drive.