The experimental environment is shown in Table 3 and we used the 2.4.4 distribution of the Linux kernel. To evaluate the effect of our coloring method, we ran the following two benchmarks on an 8-way IA server machine: WebBench 3.0 and Chat micro benchmark [7,8,9]. A lot of processes are created when these benchmarks are running. In particular, since the run queue length of Chat is longer than that of WebBench, lock contention is expected to be higher. We can expect that the coloring effect in Chat micro benchmark shows larger scalability than in WebBench.
In the case of WebBench, 256 maximum client-threads running on 28 standard PCs, simultaneously send requests to the server machine. During our experimentation, 256 httpd processes are always runnable on the server machine.
The Chat micro benchmark is a stand-alone type benchmark, and no client machine is used. This benchmark simulates chat rooms with multiple users exchanging messages using TCP sockets. Each chat room consists of 20 users, and each user broadcasts a number of 100 byte messages in the room. To handle message exchange, one user program creates four threads. Thus 80 threads are created per room. The characteristic parameters of the Chat micro benchmark are the number of rooms and the number of messages per user. We choose 30 rooms and 300 messages per user as parameters, so that 2400 processes (threads) are generated on a system during experimentation.
In this research, we also evaluated the multi-queue [8] (we called MQ) scheduler proposed by IBM, which is an alternative to vanilla scheduler. MQ scheduler separates the run queue and the run queue lock for each CPU in the system.
8