Spilling the Beans on Java Accelerators

Java has finally made it into the portable and wireless market due to Sun's Java-2 MicroEdition (J2ME). The traditional approach to execute Java is to implement the entire Java Virtual Machine in software. To meet the performance and power demands of these battery-powered applications, vendors are providing hardware accelerators to more efficiently execute Java byte codes. The Java accelerators range from extensions to the native processor's decoding hardware to standalone coprocessors that run in parallel with a host CPU.

This report focuses on the different categories of hardware-based Java accelerators including ARM's Jazelle, Aurora VLSI's DeCaf, inSilicon's JVXtreme, Nazomi's JSTAR, Parthus Technologies' MachStream, and Zucotto Wireless' Xpresso. These vendors claim that their accelerators produce an average of five to ten times increase over the speed of a software-only approach, but in reality each product has its own set of pros and cons.

• Executive Summary
  
• Methodology
  
• Nomenclature
  
• Introduction
  
• Java Hardware Interpreters
  
  
  
  • Nazomi’s JSTAR
    
  • Jazelle Pumps Up ARM’s Java
    
  
  
  
• A Memory-Mapped Coprocessor
  
  
  
  • InSilicon’s JVXtreme
    
  • Byte Codes Percolate Through Pipeline
    
  • Please Pass the Beans
    
  
  
  
• Heterogeneous Multiprocessing for Java Applications
  
  
  
  • How DeCaf Stacks Up
    
  • Is the Xpresso a DeCaf?
    
  • Built-in Java Optimizations
    
  • Branches: To Take or Not to Take
    
  • Xpresso and Host Run Simultaneously
    
  
  
  
• Hardware-Based JIT Compilation
  
  
  
  • Pouring Java Through the MachStream Engine
    
  • Brewing the Optimal Code
    
  • MachStream Deliverables
    
  
  
• Selecting the Best Brew
  
  
  
  • EEMBC Provides Standardized Benchmark Format
    
  • Theoretical Performance Analysis
    
  • How the Java Machines Stack Up
    
  • The Verdict
    
  
  
• Appendix
  

• Table 1. ARM’s mode
  
• Table 2. Quantitative comparison of Java acceleration products
  
• Table 3. Detailed comparison of byte codes supported by each Java accelerator or processor
  

• Figure 1. Nazomi’s JSTAR
  
• Figure 2. Modifications to the implementation of JVM
  
• Figure 3. JVM
  
• Figure 4. DeCaf’s pipeline
  
• Figure 5. Xpresso 100
  
• Figure 6. Simplified version of Xpresso processor’s pipeline
  
• Figure 7. Byte-code flow through MachStream
  
• Figure 8. MachStream’s division into three separate clock domains