Workloads

Table 3: Summary of fstress workloads used in the experiments.

workload	file popularities	file sizes	dir sizes	I/O accesses
SPECsfs97	random 10%	1 KB - 1 MB	large (thousands)	random r/w
Web server	Zipf ( $0.6<\alpha<0.9$)	long-tail (avg 10.5 KB)	small (dozens)	sequential reads
DB_TP	few files	large (GB - TB)	small	random r/w
Mail	Zipf ( $\alpha = 1.3$)	long-tail (avg 4.7 KB)	large (500+)	seq r, append w

We use Fstress to generate $\vec{W}$ corresponding to four workloads as summarized in Table 3. A brief summary follows. Further details are in [1].

• SPECsfs97: The Standard Performance Evaluation Corporation introduced their System File Server benchmark (SPECsfs) [6] in 1992, derived from the earlier self-scaling LADDIS benchmark [15]. A recent (2001) revision corrected several defects identified in the earlier version [11].

• Web server: Several efforts (e.g., [2]) attempt to identify durable characterizations of the Web. We derive the distributions for various parameters and the operation mix from the previous published studies (e.g., [19,8,18,9,2]).

• DB_TP: We model our database workload after TPC-C [7], reading and writing within a few large files in a 2:1 ratio. I/O access patterns are random, with some short (256 KB) sequential asynchronous writes with commit (fsync) to mimic batch log writes.

• Mail: Electronic mail servers frequently handle many small files, one file per users' mailbox. Servers append incoming messages, and sequentially read the mailbox file for retrieval. Some users or servers truncate mailboxes after reading. The workload model follows that proposed by Saito et al. [20].