Shen Meiming

Quasi-asynchronous migration: a novel migration protocol for PVM tasks

Process migration is a desirable and useful facility for the Network of Workstations (NOW). This paper presents a novel migration protocol called quasi-asynchronous migration for PVM tasks, which allows non-migrating tasks to execute during most of the time of migration. Message flushing and message delaying are the key mechanisms used in quasi-asynchronous migration. Because the protocol is implemented on top of PVM and at user-level, it is both transparent to users and portable. Both the analysis of the migration protocol and the experimental results show that quasi-asynchronous migration has lower overhead than other migration protocols.Process migration is a desirable and useful facility for the Network of Workstations (NOW). This paper presents a novel migration protocol called quasi-asynchronous migration for PVM tasks, which allows non-migrating tasks to execute during most of the time of migration. Message flushing and message delaying are the key mechanisms used in quasi-asynchronous migration. Because the protocol is implemented on top of PVM and at user-level, it is both transparent to users and portable. Both the analysis of the migration protocol and the experimental results show that quasi-asynchronous migration has lower overhead than other migration protocols.

The compiler for supporting multithreading in cyclic register windows

This paper presents the compilation technique to support the P&M (pipelining and multithreading) multithreaded program execution model on the SPARC processor, which has cyclic register windows. The program execution model improves the performance by running multiple user-level threads on each processor and filling the idle cycles that would otherwise be wasted waiting for messages. The integrated compilation paradigm is described systematically. The generated code is efficient, which obtains rapid context switching without sacrificing the benefit of fast procedure call.

Research on data replica distribution pattern for web server clusters

Today more and more WWW applications ask for basically non-stop services around the clock. Finding a good highly available scheme for a cluster web server becomes very important. Web server clusters with different data replica distribution pattern have different data availability level. Therefore our research focuses on data replica distribution pattern. After making theoretical analysis on data replica distribution patterns, an optimized one is put forward in this paper.

On communication patterns of parallel objects

Object-oriented programming is proving to be quite useful in many areas which have not been served well by more conventional approaches. Object-oriented systems show great potential for exploiting concurrence/parallelism. But many object-oriented languages currently in use are sequential. How to implement object-oriented languages on parallel systems is still a promising research topic. A parallel object-oriented system is a single-user system which is implemented on multi-processors. Objects can be active concurrently in one processor or do things on different processors in parallel. There are two communication patterns: synchronous and asynchronous communication. We put emphasis on communication patterns of parallel objects. We also discuss several ways to achieve performance in object-oriented languages and how to implement communication patterns on transputer-based multicomputer systems.<<ETX>>

A distributed naming mechanism in scalable cluster file system

The authors present a novel distributed naming mechanism (DNM) that is fit for network file system design. With the help of DNM, we build a new distributed file system paradigm: cluster file system. In our system, DNM not only offers the desired features of high availability, high scalability and I/O load balancing, but also spans a single volume on multiple servers to construct a real distributed file system. We propose the DNM structure which relies on a global replicated table and a local information database. To improve the performance of name translation, a tree-based name cache is introduced. A prototype of the name mechanism has been implemented in our cluster file system.

The approach to support the P&M multithreaded program execution model on workstation cluster

This paper presents the approach to support the P&M (Pipelining and Multithreading) program execution model on workstation cluster. The inter-node pipeline of communication and computation on workstation cluster is aimed to be constructed. The multithreaded program execution model employs both the incoming message handler and outgoing message handler to draw the distinction between real computation and communication. Low overhead communication is supported by the proposed communication protocol. The logical organization of one P&M node is developed which aims to support communication and computation by separate units. We further describe the P&M C language and the compilation technique. Emulation of the SPARC workstation based P&M platform convinces us that improved overall performance can be achieved, and more demanding applications can be feasibly solved by the P and M approach.

ON-LINE DEBUGGING OF PARALLEL PROGRAMS

Nondeterminacy is a main obstacle of parallel debugging. Current debugging strategies either detect non-dcterminacy separately, or control the execution in a two-phase manner. Here we present a strategy called “state freezing”. Based on the strategy, one-phase (we call one-line) debugging of parallel programs is available. Basic algorithms and an example are given in the paper.

Compiling CIL rewriting language for multiprocessors

The high-level Compiler Intermediate Language CIL is a general-purpose description language of parallel graph rewriting computational model intended for parallel implementation of declarative languages on multiprocessor systems. In this paper, we first outline a new Hybrid Execution Model(HEM) and corresponding parallel abstract machine PAM/TGR based on Extended parallel Graph Rewriting Computational Model EGRCM for implementing CIL language on distributed memory multiprocessor systems. Then we focus on the compiling CIL language with various optimizing techniques such as pattern matching, rule indexing, node ordering and compile-time partial scheduling. The experimental results on a 16-node Transputer Array demonstrates the effectiveness of our model and strategies.

Development of the parallel inference machine RAP/LOP-WAM and its optimized parallel compiler

A brief overview is presented of the parallel abstract machine developed for the RAP/LOP parallel execution model. A description is also given of its optimized parallel compiler. The main features of the machine are: (i) the OR-forest description is used to describe the search space of a given problem, not only describing OR- and AND- parallelism explicitly, but also avoiding a class of redundant computations; (ii) coarse-grain parallelism is supported by the granularity-based scheduling policy; (iii) procedure-level and clause-level analysis at compile-time and dynamic simple run-time checks are used to identify independent goals of the body of a clause; and (iv) several optimization and implementation techniques such as improved indexing mechanism and code space reduction are used to increase the machines efficiency significantly.<<ETX>>