This paper presents communication strategies for achieving efficient parallel and distributed Java applications on clusters with high-speed interconnects. Communication performance is critical for the overall cluster performance. Previous efforts at obtaining efficient Java communications have a limited applicability on high-speed interconnects as they are focused on high level APIs like RMI, ignoring the particularities of these systems and their native high performance communication protocols. By relying on a custom Java socket implementation higher degrees of performance can be achieved exploiting high-speed interconnect facilities. Several protocol definitions are presented, looking for obtaining high performance Java communications. Moreover, the quality of the protocol implementations and their design decisions has been thoroughly evaluated on a Scalable Coherent Interface (SCI) and Gigabit Ethernet (GbE) testbed cluster. The results of this analysis have demonstrated that these Java protocols obtain similar results to native communications. 1

Remote Method Invocation (RMI), a mechanism to access remote objects in Java-based distributed applications, uses network communication for each method invocation. Consequently, using RMI in a wide-area environment can cause poor application performance. One solution to improve performance is to cache the objects such that network communication is not necessary for each method invocation. In this paper, we present mechanisms to transparently add object caching to RMI. These mechanisms are compatible with existing RMI applications and use an event-based model to support different consistency policies. The mechanisms also include the ability to adaptively select the consistency policy for an object based on its usage pattern. A novel feature of our mechanisms is the use of a ‘reduced object’, which is a partial representation of the RMI object. We experimentally evaluate and demonstrate the benefits of our

Program Generation (PG) is about writing programs that write programs. A program generator composes various pieces of code to construct a new program. When employed at runtime, PG can produce an efficient version of a program by specializing it according to inputs that become available at runtime. PG has been used in a wide range of applications to improve program efficiency and modularity as well as programmer productivity. There are two major problems associated with PG: (1) Program generation has its own cost, which may cause a performance loss even though PG is intended for performance gain. This is especially important for runtime program generation. (2) Compilability guarantees about the generated program are poor; the generator may produce a type-incorrect program. In this dissertation we focus on these two problems. We provide three techniques that address the first problem. First, we show that just-in-time generation can successfully reduce the cost of generation by avoiding unnecessary program generation. We do this by means of an experiment in the context of marshalling in Java, where we generate specialized object marshallers based on object types. Just-in-time generation improved the speedup from 1.22 to 3.16. Second, we apply source-level transformations to optimize

acceptée sur proposition du jury: Prof. A. Wegmann, président du jury Prof. M. Odersky, directeur de thèse