As we discussed in more length in Subsection 5.1, Postmark load is high and we choose to benchmark also a case of medium load in the system by compiling simultaneously two and four Linux kernels. Under high load, as it is the case of Postmark, IO requests wait mostly at the device driver, because the number of outstanding IO requests is higher than the maximum disk queue depth. Hence disk level queuing is effective only for a small portion of the outstanding IO requests and optimization effectiveness at the device driver becomes more important than queuing at the disk drive level.
Under medium loads, as it is the case of the Linux kernel compilation, which we described in Subsection 5.1, the impact of disk queuing in the overall system performance is higher than under high loads. This holds in particular for cases when the optimization at the device driver level (or any other level in the IO path) is not as effective as it could be (as it is the case of the No-Op scheduler).
In Figure 17, we show the effectiveness of disk queuing when compiling 2 and 4 Linux kernels simultaneously. Note that No-Op benefits the most from disk queuing and the result is that the overall system performance (measured in builds per seconds) is very close for all four device driver disk schedulers when disk queue depth is 16 or higher. Hence, we conclude that disk level queuing and scheduling is effective and does close any performance gaps in the IO request reordering from the higher levels in the IO path.