Effect of Technology Trends

 

 

Table 2:   Performance gap between update-in-place and virtual-logging widens as disks and processors improve.

The performance improvement provided by eager writing seen so far is not as great as the analytical models might suggest. To see why this is the case, we now provide a more detailed breakdown of the benchmark times reported in the last section. We also examine the impact of technology trends.

We repeat the UFS experiment of the last section on three different platforms under the same disk utilization (80%). Table 2 shows the result. The first column shows the speed-up of virtual logging compared to update-in-place on a SPARCstation-10 and HP disk combination. Next, we replace the older HP disk with the new Seagate disk. In the third run, we replace the older SPARCstation-10 with the newer UltraSPARC-170. We see that the performance gap widens to almost an order of magnitude.

 

Figure 9:   Breaking down the total latency into SCSI overhead, transfer time, time required to locate free sectors, and other processing time. Update-in-place performance (left bars) becomes increasingly dominated by mechanical delays while virtual logging (right bars) achieves a balance between processor and disk improvements.

Figure 9 reveals the underlying source of this performance difference by providing the detailed breakdown of the latencies. The ``SCSI overhead'' component is the time that the disk processor spends processing each SCSI command. The ``transfer'' component is the time it takes to move the bits to or from the media after the head has been positioned over the target sector. The component labeled as ``locate sectors'' is the amount of time the disk spends positioning the disk head. It includes seek, rotation, and head switch times. The ``other'' component includes the operating system processing overhead, which includes the time to run the virtual log algorithm because the disk simulator is part of the host kernel. The time consumed by simulating the disk mechanism itself is less than 5% of this component.

We see that the mechanical delay becomes a dominant factor of update-in-place latency. We also see that eager writing has indeed succeeded in significantly reducing the disk head positioning times. The overall performance improvement on the older disk or on the older host, however, is low due to the high overheads. After we replace the older disk, the performance of virtual logging becomes host limited as the ``other'' component dominates. After we replace the older host, however, the latency components again become more balanced. This indicates that the virtual log is able to ride the impressive disk bandwidth growth, achieving a balance between processor and disk improvements.