Disk Drive Level

The disk drive is the last component in the IO path that we analyze in this paper. With the advances in chip design, modern disks have more CPU and more memory available than before. Although various optimization techniques such as request merging are best suited for the upper levels of the IO path such as the file system and the device driver, the disk itself does offer various optimization opportunities, mostly related to caching of data and scheduling of requests. We stress that disk scheduling at the disk level is the most effective IO request scheduling in the IO path [22], because it uses information on head position that is available only at the disk drive level. In this section, we focus on evaluating disk level scheduling by analyzing the effectiveness of disk level queuing.

We use three different disks in our evaluation manufactured by Seagate Technology. The main differences between them are the linear density, rotational speed, and capacity, which determine the average head positioning time and the associated overhead for each disk request. Details on the three disks used in our evaluation are given in Table 3. Note that the linear density of the 300 GB disk is approximately 24% and 64% higher than the linear density of the 146 GB disk and the 18 GB disk, respectively, while its average seek time is higher by approximately 30%. Although seek time is expected to improve if disk tracks are closer together, in the case of the 300 GB drive, the platters are larger to accommodate the large capacity and consequently the arm itself becomes heavier, which results in longer seek times for the 300 GB disk than the 18 GB and 146 GB disks.

Table 3: Specifications for the three Seagate disks.

	ST318453LC	ST3146854LC	ST3300007LC
Capacity	18 GB	146 GB	300 GB
RPM	15,000	15,000	10,000
Platters	1	4	4
Linear density	64K TPI	85K TPI	105K TPI
Avg seek time	3.6/4 ms	3.4/4 ms	4.7/5.3 ms
Cache	8 MB	8 MB	8MB

In Figure 13, we show the Postmark throughput for the three disks of Table 3 and the SS and LL workloads of Table 2. The file system is changed while the disk scheduling is set to Anticipatory. The highest application throughput is achieved for the 146 GB disk, with the 18 GB to a close second. The higher average seek times and the slower rotation speed cause the 300 GB disk drive, although with the highest linear density, to achieve the lowest application throughput among the three disk drives.

Figure 13: Postmark throughput for all disks of Table 3, two of the workloads of Table 2, the four file systems, anticipatory scheduling, and disk queue depth of 1.

Figure 14 is similar to Figure 13, but now the file system is fixed to ReiserFS and the Postmark performance is measured under four disk scheduling algorithms. As pointed out above, the 146 GB disk is the one performing best under all device driver disk scheduling algorithms. Note that the relative gap between No-Op and the other seek-based algorithms is smaller for the newer disks (i.e., 146 GB and 300 GB) than the older one (i.e., 18 GB).

Figure 14: Postmark throughput for all disks of Table 3, two of the workloads of Table 2, the four disk scheduling algorithms, ReiserFS, and disk queue depth of 1.