Operating System for Parallel Computing

When available computers are incapable of providing a sufficient computing power for solving the arising tasks, while purchasing new and more powerful equipment is economically infeasible, then the only way to gain additional computing power is to parallel the computational process between several interconnected computers. The idea of parallel distributed computing has been studied in cybernetics since early sixties. However, until recently parallelism was primarily employed within the framework of a single computer only. For example, it formed the basis for extending the classical von Neumann architecture to superscalar processors and later – to explicit parallel instruction computing (EPIC) architectures [1]. The majority of computational tasks can be implemented in concurrent manner. The highest level of parallelism is achieved when the task can be represented by a set of loosely coupled sequential processes that have only few synchronization points. Modern operating systems support parallel execution of processes on multiprocessor and uniprocessor computers (the latter form of parallelism is known as pseudo-parallelism). For this purpose an operating system provides process synchronization and communication facilities. However, a process is limited to a single computational node. In order to implement a parallel algorithm in a multicomputer environment application programmer has to involve additional libraries and tools for organizing synchronization and data exchange between remote processes. Obviously, this task can be more effectively addressed by an operating system. Recently, we started the work on a new operating system aiming to support transparent execution of parallel algorithms in a distributed environment. This system will provide the distribution mechanisms, which will allow a developer to migrate his parallel application from a single computer to a network with minimal efforts. The new system will support dynamic balancing of a computational payload among network nodes. For this purpose it will provide each process with execution environment which will not be bound to a specific machine. Thus, a process started in one node can be transparently transferred to another node at any time. This procedure known as process migration will form the basis for building the distributed execution environment. Currently there exist a number of approaches to carrying out computations in distributed environments. Most of them rely on special parallel programming libraries that provide application developers with message-based distributed communication and synchronization facilities. Two well-known examples of such libraries are Message Passing Interface (MPI) [2] and Parallel Virtual Machine (PVM) [3]. This approach allows easily building a distributed programming environment on the basis of a network of computers running conventional operating systems like UNIX or Windows NT. However, parallel programming libraries do not support process migration between network nodes. Besides, these libraries are based on highlevel operating system services, which makes it easy to implement them but negatively affects their performance. The most important drawback inherent to MPI, PVM and other programming platforms of this kind is that each of them imposes its own specific process interaction mechanisms. As a consequence, adaptation of a parallel program for execution in a distributed environment requires significant efforts for porting it to a new set of communication primitives. An interesting alternative approach was taken by MOSIX [4] project aimed to build an extension of the Linux kernel, which would support distributed computing by presenting processes with a single execution environment within a system of several interconnected computers. MOSIX implements efficient process placement and migration policies and provides a distributed shared memory facility.