4.1 Methods

To verify the cache coloring improvement, we chose a web server program as a scheduler intensive benchmark, and evaluated the performance improvement. In the experiment, Apache 1.3.19 is used as the web server, and two Linux kernels are used for comparison: vanilla Linux 2.4.4 kernel and modified kernel with using 32 colors for the task structure. The web clients ran WebBench 3.0 [6] to generate web requests for the server. The specification of the server machine is shown in Table 1. We used three different sizes of L2 cache for evaluating the coloring effects, and all of them have four-way set associative caches. During the benchmark execution, 256 client threads were running on 28 standard PCs and they simultaneously sent requests to the server machine. On the server-side, 256 to a maximum of 1024 server threads (httpd) ¹were runnable and were linked on the run queue of the server system.

Table 1: The specification of the server machine

CPU	Pentium Pro 200MHz $\times$ 4
Memory	640MB
L2 cache	256KB, 512KB, 1024KB
	4 way set associative
NIC	Intel EtherExpress Pro/100 $\times$ 4