Real-world Traces

Four real-world file system traces we used in our simulation. We got two sets of real-life traces from two different universities to validate our results. Two of them came from University of California, Berkeley, called INS and RES [14]. INS came from a collection from a group consisting of 20 machines located in labs for undergraduate classes. RES was attained from 13 desktop machines of a research group. INS and RES were recorded over 112 days from September 1996 to December 1996. Both traces came from their clusters running HP-UX 9.05. The other set of two traces, from University of Kentucky, contain all disk activities on two SunOS 4.1 machines during ten days for Sitar trace and seven days for Harp trace[6]. Sitar trace represents an office environment while Harp reflects common program development activities. More specifically, Sitar trace is a collection of file accesses by graduate students and professors doing work such as emailing, compiling programs, running LaTeX, editing files, and so on. Harp trace shows a collaboration of two graduate students working on a single multimedia application. Because Sitar and Harp have a small amount data, we use the small disk model with these two real-world traces. Notice in the experiments, we expand Sitar and Harp by appending files with same access pattern in original traces but with different file names in order to explore the system behavior under different disk utilizations. For large traces with more than 10GB data traffic, we do not use this procedure.

These real-world traces are described in more detail in Table  1.

Table 1: Four real-world traces and Four synthetic traces

Features	INS	RES	SITAR	HARP	Uniform	Hot-cold	TPC-D	SmallFS
Data read(MB)	94619	52743	213	520	8000	8000	8000	8000
Data write(MB)	16804	14105	183	249	8000	8000	4000	8000
Read:Write ratio	5.6	3.7	1.16	2.08	1.0	1.0	2.0	1.0
Reads(thousands)	71849	9433	57	26	800	800	800	4000
Writes(thousands)	4650	2216	49	12	800	800	400	4000
File Size($<$ 16KB)	82.8%	63%	80%	73	80%	80%	0%	95 %
File Size(16KB-1MB)	16.9%	36.5%	19.96%	26.98%	19.95%	19.95%	0%	5%
File Size(1MB+)	0.2 %	0.5 %	0.04 %	0.02 %	0.05%	0.05%	100%	0%