Distributed Memo: A Heterogeneously Distributed and Parallel Software Development Environment

Abstract

Abstract

Java on networks of workstations (JavaNOW): a parallel computing framework inspired by Linda and the Message Passing Interface (MPI)

“…While a detailed description of the implicit synchronization mechanism is beyond the scope of this paper, we note that it is based on an abstraction called a shared directory of unordered queues, which has been addressed in detail in our Memo papers [4,11], which was one of two predecessors (the other being Enhanced Actors [12]) to the JavaNOW system.…”

“…This placement control can be achieved to a large extent because the Java language Object class provides a default definition of a hashcode() method, which guarantees that any intrinsic or user-defined class that is used to create objects will produce a hash value, albeit not a well-conditioned hash value. Precise placement control can be very easily achieved in JavaNOW due to a design decision that was made to clearly isolate the key from the rest of the tuple, which was first done in previous work published by one of the authors on the Memo system [17] [18]. The way precise placement control is achieved (discussed in detail later) is to provide a custom hashcode() function that can be used to compute (mod the number of contexts) the precise destination for an entity.…”

“…All put and get operations that manipulate the same Entity are guaranteed to be mutually exclusive freeing the programmer from concerns related to the synchronization of the EntitySpace. While a detailed description of the implicit synchronization mechanism is beyond the scope of this paper, we note that it is based on an abstraction called a shared directory of unordered queues (a shared data structure that guarantees mutual exclusion and has entries with mutually-exclusive access), which was discussed in two of our earlier papers on Memo [17] and Enhanced Actors [62].…”

“…JavaNOW is a pure Java-based system that can execute on any architecture with Java Virtual Machine (JVM) support without recompilation. This is in contrast to other more traditional network-based frameworks, such as MPI [1], PVM [2], Network Linda System [3], and Memo [4] that require architecture-specific binary libraries and executables to be installed on each machine on which the virtual parallel computer will execute. In addition to traditional systems, there are a number of competing pure-Java frameworks (most notably Sun JavaSpaces TM ) that have also been inspired by the Linda tuple space model; however, these systems have not been designed with highperformance distributed computing in mind and are missing some key features that would enable higher performance, including active tuple evaluation and collective communication facilities.…”