Real Workloads

The first two synthetic benchmarks focused on file creation in empty or partially-full file systems, demonstrating some of the benefits of range writes. We now simulate the performance of an application-level workload. Specifically, we focus on three workloads: untar, which unpacks the Linux source tree, PostMark [19], which simulates the workload of an email server, and the modified Andrew Benchmark [15], which emulates typical user behavior. Table 2 presents the results.

Table: File System Workloads. Each row plots the performance (in seconds) of a simulated workload. In the left column, results represent the time taken to run the workload on our simulated standard ext2, whereas on the right, the time to run the workload on ext2 with range writes is presented. Three workloads are employed: untar, which unpacks the Linux source tree; PostMark, which emulates the workload of an email server (by creating, accessing, and deleting files), using its default settings; and the modified Andrew benchmark, which emulates typical user behavior. The simulations were driven by file-system-level traces of the given workloads which were then played back against our simulated file system.

	Traditional ext2	with Range Writes
Untar	143.0	123.1
PostMark	29.9	22.2
Andrew	23.2	23.4

0.9

We make the following two observations. First, for workloads that have significant write components (untar, PostMark), range writes boost performance (a 16% speedup for untar and roughly 35% for PostMark). Second, for workloads that are less I/O intensive (Andrew), range writes do not make much difference.