Year: 2000
Type: conference-paper
Abstract: Although Java promises platform portability amongst diverses et of systems, for most Java platforms today, it is not clear if they are appropriate for high-performance numerical computing.In fact, most previous attempts at utilizing Java for HPC sacrificed Java's portability, or did not achieve necessary performance required for HPC.Instead, we propose an alternative methodology based on Downloadale Self-tuning Library, and constructed an experimental prototype called AJaPACK, which is a portable and highperformance parallel BLAS library for Java which "tunes" itself to the environment to which it is installed upon.Once AJaPACK is downloaded and executed, the Java version of ATLAS (ATLAS for Java) and the parallelized version of JLAPACK combine to achieve optimized pure Java execution for the given environment.Benchmarks have shown that AJaPACK achieves approximately 1/2 to 1/5 of the speed of optimized C-ATLAS and vendor supplied BLAS libraries, and with portable parallelization in SMP environnmnts, achieves superior performance to single-threaded Cbased native libraries.This is an order of magnitude superior w.r.t, performance compared to previous pure Java BLAS libraries, and opens up further possibilities of employing Java in HPC settings, but still shows that J I T compilers with optimizations expecting numerical code highly-tuned at the source-or bytecode level would be highly desirable.
Institution: Tokyo Institute of Technology
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
Funder: University of Tokyo
OpenAlex ID: W1982123697