Trace Workload

We now present preemptibility results using IO traces obtained from a Linux system. IO traces were obtained from three applications. The first trace (DV15) was obtained when the XTREAM multimedia system [6] was servicing $15$ simultaneous video clients using the FCFS disk scheduler. The second trace (Elevator15) was obtained using the similar setup where XTREAM let Linux elevator scheduler handle concurrent disk IOs. The third was a disk trace of the TPC-C database benchmark with 20 warehouses obtained from [15]. Trace summary is presented in Table 1.

Table 1: Trace summary.

Trace	Number of req.	Avg. req. size	Max. block number
		$[blocks]$	$[blocks]$
DV15	10800	128.7	28442272
Elevator15	10180	127.6	28429968
TPC	1376482	126.5	8005312

Figure 14: Improvement in the expected waiting time (using disk traces).

Figure 15: Effect on the achieved disk throughput (using disk traces).

Figures 14 and 15 show the expected waiting time and disk throughput for these trace experiments. The expected waiting time was reduced by as much as $65\%$ (Figure 14) with less than $10\%$ (Figure 15) loss in disk throughput for all traces. (Elevator15 had smaller throughput than DV15 because several processes were accessing the disk concurrently, which increased the total number of seeks.)