Using the CPU timer

Introduction

It is possible to use the native JVM profiling API to obtain more accurate times from JBench. The granularity isn't usually affected (eg on Windows 2000 the granularity is still 10ms) but the actual CPU time spent executing code in the main JBench thread is reported, rather than the "wall-clock" time which is affected by how heavily loaded your computer is. I am indebted to JavaWorld and in particular the Java Tip 92 article for the basis of the code used.

Installing the native library

This interface requires a native library, which must be installed. The source for the library and pre-built binaries for Linux (i386), Solaris (sparc) and Windows are provided in the native directory of both the source and binary distributions of JBench. Generally, under Windows you should place cputimer.dll in your jre/bin directory, under Linux you should place libcputimer.so in your java/lib/i386 directory, and under Solaris you should place libcputimer.so in your java/lib/sparc directory. Different JVMs may differ in how they pick up native libraries, however, so consult your JVM documentation when in doubt.

Configuring JBench to use the CPU timer

In your JBench configuration file, set jbench.timer=cpu to use the native library. If you're running on Linux and see absurdly small times being reported, also set jbench.timer.scale=1000 - as far as I can see, there's a bug in some HotSpot JVMPI implementations which give values in microseconds instead of nanoseconds.

Running JBench with the native library

On the command line, you should use -Xruncputimer as a command line option to the VM, eg:
java -Xruncputimer -jar JBench.jar jbench.properties
This will initialise the library properly before it is used.

JBench Manual