William F. Appelbe

Integrating tools for debugging and developing multitasking programs

For the foreseeable future, large numbers of scientific applications programs will continue to be written in Fortran dialects, and the widespread use of multiprocessors will not result in any change in the status quo. Indeed, the popularity of shared memory multiprocessors can be traced to the ease with which sequential programs and programming languages can be adapted to exploit the hardware’s parallelism. Such ease of adaptation is unfortunately restricted to coding alone multitasking Fortran is notoriously difficult to debug and maintain using present software technology. The major tools in use for sequential program development optimizing compilers and source level debuggers, can be extended to support a paradigm in which the user continues to program in a sequential language, and the compiler introduces task-level parallelism. There are three major limitations in this approach

The architecture of Ra: a kernel for Clouds

Ra is a native, minimal kernel for the Clouds distributed operating system. Ra is a successor to the prototype Clouds kernel and reflects lessons learned from the earlier implementation effort. Ra supports the same object-thread model as the original Clouds kernel as a special case and introduces extensibility as a major goal. Ra provides three primitives, namely segments, virtual spaces, and lightweight processes called isibas, which can be composed in various ways to construct components of the Clouds operating system. The architecture and organization of Ra and details of its implementation are described. The authors describe the implementation of several Clouds components such as objects and threads using Ra primitives to demonstrate the versatility and power of the Ra kernel. Ra supports plug-in system objects that allow system services to be introduced and removed dynamically. Ra has been implemented in C++ and is running on Sun-3 workstations.<<ETX>>

A formal protection model of security in centralized, parallel, and distributed systems

One way to show that a system is not secure is to demonstrate that a malicious or mistake-prone user or program can break security by causing the system to reach a nonsecure state. A fundamental aspect of a security model is a proof that validates that every state reachable from a secure initial state is secure. A sequential security model assumes that every command that acts as a state transition executes sequentially, while a concurrent security model assumes that multiple commands execute concurrently. This paper presents a security model called the Centralized-Parallel-Distributed model (CPD model) that defines security for logically, or physically centralized, parallel, and distributed systems. The purpose of the CPD model is to define concurrency conditions that guarentee that a concurrent system cannot reach a state in which privileges are configured in a nonsecure manner. As an example, the conditions are used to construct a representation of a distributed system.

Determining Transformation Sequences for Loop Parallelization

Considerable research on loop parallelization for shared memory multiprocessors has focused upon developing transformations for removing loop-carried dependences. In many loops, more than one such transformation is required, and hence the choice of transformations and the order in which they are applied is critical.

Program Transformation for Locality Using Affinity Regions

Affinity regions ensure that a shared processor schedule, mapping loop iterations to processors, is used in consecutive parallel loop nests. Using affinity regions can improve locality without affecting parallelism.

A New Algorithm for Global Optimization for Parallelism and Locality

Converting sequential programs to execute on parallel computers is difficult because of the need to globally optimize for both parallelism and data locality. The choice of which loop nests to parallelize, and how, drastically affects data locality. Similarly, data distribution directives, such as DISTRIBUTE in High Performance Fortran (HPF), affects available parallelism and locality. What is needed is a systematic approach to converting programs to parallel form, based upon analysis that identifies opportunities for both parallelism and locality in one representation.

Hoisting Branch Conditions - Improving Super-Scalar Processor Performance

The performance and hardware complexity of super-scalar architectures is hindered by conditional branch instructions. When conditional branches are encountered in a program, the instruction fetch unit must rapidly predict the branch predicate and begin speculatively fetching instructions with no loss of instruction throughput. Speculative execution has a high hardware cost, is limited by dynamic branch prediction accuracies, and does not scale well for increasingly super-scalar architectures.

The hierarchical model of distributed system security

A description is given of the hierarchical model (HM), an access matrix-based model used to define nondisclosure in distributed multilevel secure applications such as secure file systems, secure switches, and secure upgrade downgrade facilities. The HM explicitly encodes access rights, synchronization primitives, and indirection in its state matrix. Serializability of concurrent commands is formally defined in terms of the HM syntactic model of computation. HM serializability conditions are independent of the semantic security predicate. Finally, an example that illustrates the HM is presented.<<ETX>>