Old CACAO Home Page

Out-dated page

CACAO is 64 bit just-in-time (JIT) compiler for Java. It translates Java byte code on demand into native code for the ALPHA processor. During this translation process the stack oriented Java byte code is transformed into a register oriented intermediate code. Local variables and stack locations are replaced by pseudo registers eliminating the 32 bit restriction on address types. A fast register allocation algorithm is applied to map the pseudo registers to machine registers. During code generation, field offsets are computed for proper alignment on 64 bit architectures. Even though the CACAO system has to incur loading and compilation time, it executes Java programs up to 85 times faster than the JDK interpreter, up to 7 times faster than the kaffe JIT compiler. It is slightly slower than equivalent C programs compiled at the highest optimization level. The design and implementation of CACAO (old JIT) is described in an article (44686 bytes) for the PPoPP'97 Workshop on Java for Science and Engineering Computation and in an article (29896 bytes) which will appear at PACT'98 (new JIT). The implementation of synchronization and exception handling is described in an article (63472 bytes) for the ACM 1998 Workshop on Java for High-Performance Network Computing.

CACAO is a replacement for the interpreter java of the JDK from SUN. How to invoke CACAO is described in the CACAO manual page.

The beta version of CACAO 0.2b with the new faster JIT and support for multithreading is available for the Alpha processor with Digital Unix. The Linux version will follow very soon. Version 0.1 supports the Alpha processor under Linux and Digital Unix. It can be used with the class library of the JDK 1.0.2. CACAO version 0.1 does not support AWT, it only supports java/lang, java/io and java/util. Class archives in zip format are not supported. CACAO is distributed free of charge. Currently only binary versions are distributed. The source will be available in the future (under GNU licence requirements). The CACAO installation page describes how to download and install CACAO.

CACAO is a research prototype under development. The next version will include a new garbage collector and will support the net library and additional processors (planned for christmas 98). Code generators for the MIPS, PowerPC and Sparc processors are under development. We also started working on the integration of the BISS-AWT libarary. We will increase the speed of the generated code by instruction scheduling and the speed of the compiler by using faster algorithms. It is quite likely that CACAO contains bugs. Please send bug reports to cacao@complang.tuwien.ac.at. If you want to participate in the development of CACAO or have urgent questions send mail to andi@complang.tuwien.ac.at.

To evaluate the performance of CACAO we compared it with Sun's JDK and with kaffe version 0.8 and the newest version of the Digital JIT 1.1.1 of November 6th 1997. We also compared it with two other JavaVM to native code compilers for the SPARC processor:

Both systems are not available for the Alpha and CACAO currently supports only the Alpha. So we compared the all systems with the JDK interpreter (1.0.2 OSF port), assuming that a SPARC processor and an Alpha processor are similar and also the implementation of the JDK interpreter is similar.

The benchmark programs and the run time data for Guava and Toba are taken from the Toba homepage. We tested CACAO with exactly the same programs and the same data. JavaLex is a scanner generator, javac is the Java compiler from the JDK compiling the Toba sources, espresso is another Java compiler, Toba is a JavaVM to C compiler and java_cup is a parser generator.

Table 1 gives the run time and for CACAO additionally the load time and the compile time for the benchmarks on a SparcStation 20 and an Alpha workstation with 300MHz 21064a processor. The CACAO system is in nearly all cases faster than Guava and Toba, only where the compile time is high the Toba system is faster. 

                  JavaLex    javac   espresso   Toba   java_cup

            runtime on SparcStation 20 (in seconds) 
        
JDK                176.3      45.0     24.9     59.8      8.3
GUAVA               80.8      17.8       -      36.8      6.2
TOBA                38.8      12.8      4.9     20.1      2.9

              runtime on 21064A 300MHz (in seconds)

JDK                29.8       18.5      8.7     32.1      3.5
Digital Interpr.   98.9       56.4     17.9       -       9.2
Digital JIT        84.4       47.6     14.1       -       9.8
kaffe               9.9       17.8     12.5       -       2.98
CACAO total         2.65       4.74     3.17     4.58     1.52
      load          0.18       0.61     0.38     0.18     0.21
      compile       0.39       1.21     0.91     0.55     0.81
      run           2.08       2.92     1.88     3.85     0.50

             number of compiled JavaVM instrucions

                   13412      34759    27281    14430    17489
                   
             speedup with respect to interpreter

speedup JDK/DEC-JIT 0.35       0.38     0.48      -       0.36
speedup JDK/GUAVA   2.18       2.53      -       1.62     1.34
speedup JDK/TOBA    4.54       3.51     5.08     2.98     2.86
speedup JDK/kaffe   3.01       1.04     0.7       -       1.17
speedup JDK/CACAO  11.24       3.90     2.74     7.01     2.30

                       Table 1
Table 2 compares the CACAO system with a C compiler and the newest version of the Digital JIT 1.1.1 of November 6th 1997. sieve is the well known prime number computation program, addition is a loop with a simple addition and linpack is a floating point intensive program. The option -cbnf of CACAO disables array bound checks and precise floating point exceptions. Since C does not do these checks it is fair to do the comparison without checks. The CACAO system is only a factor of 1.7 slower than C and up to 7 times faster than kaffe. 

       runtime on 21064A 300MHz (in seconds)
            
                         sieve  addition   linpack
                         
JDK                      83.2    138.76      1.6
Digital interpr.         70.3    124.4       2.0
kaffe                     9.14     4.75      0.34
Digital JIT               6.27     5.33      0.9
GCC -O3                   2.0      1.40       -
CACAO total               4.80     1.69      0.81
        run               4.58     1.42      0.58
CACAO -cbnf total         3.46     1.64      0.33
        run               3.31     1.42      0.13
        
                 relation of runtime

speedup JDK/kaffe          9.10   11.8       4.7 
speedup JDK/DEC-JIT       13.3    29.2       1.7
speedup JDK/CACAO         17.3    82.1       2.0
speedup JDK/CACAO -cbnf   24.1    85.7       4.8
speedup JDK/GCC           41.6    99.1        - 
CACAO -cbnf run/GCC        1.66    1.01       - 

                        Table 2
Last updated by Andreas Krall on October 12th 1998. Please send mail to cacao@complang.tuwien.ac.at for errors and changes.