Nancy A. Burns

The evolution of a high-performing Java virtual machine

Early JavaTM virtual machines (Jvms) possessed several significant performance bottlenecks that inhibited the speed of Java workloads. This paper presents the methodology that was used by IBM to identify and eliminate these bottlenecks for improving the performance of Java applications running on several operating system platforms. In addition, several of the key performance problems that were common to all early Java virtual machine implementations and how they were solved for IBM enhanced Jvms are described in detail. The issues discussed in this paper are focused on problems found in core Jvm components, such as object synchronization, object allocation, heap management, text rendering, run-time resolution, and Java class library methods. The results obtained from applying the described methodology and eliminating the identified performance bottlenecks increased the performance of IBM Java virtual machines by as much as four times on some workloads. The technology discussed in this paper is applicable to other Jvm implementations.

The portable inference engine: fitting significant expertise into small systems

The Portable Inference Engine (PIE) is the nucleus of an expert system that allows the segmentation of rules in order to utilize large knowledge bases in limited memory. If the expert-systems rules can be divided into segments of highly related rules with little interaction among those segments, such knowledge-base segments can then be paged in and out of memory on demand. PIE gathers information by querying the user and executing external procedures in order to conclude goals.