Proposal and Evaluation of Realization Approach for a Shared Memory System in λ Computing Environment

As users of networks such as the Internet have increased, the amount of traffic has steadily increased. Various applications that utilize images have come to be used, and the demands made on the technology that enables the high speed and large scale transmission in a network have increased. To satisfy these demands, research into optical technologies like IP over a WDM network, GMPLS, and optical packet switches have been studied and developed. However, many such technologies presuppose the existing Internet technology. That is, an IP packet is treated as a degree of granule treating information. As long as architecture based on packet switching technology is focused on, realization of high quality communication to each connection will be very difficult. New technologies such as SAN and Grid computing need to be applied to provide end users with a high speed and reliable communication pipe. In this thesis, we propose a new computing environment (which we will refer to λ computing environment) that provides a base for parallel computing among nodes distributed in the wide area. In our concept, virtual channels are provided utilizing optical networks connecting computing nodes. It can offer high-speed and reliable connection pipe among nodes, so that it is efficiently applicable to SAN (Storage Area Network) and/or Grid computing. In the environment, shared memory is constituted on a virtual ring of the photonic network. Consequently, it is not necessary to distinguish shared memory in a wide-area distributed system from a communication channel; thus high-speed data exchange between computing nodes on the ring can be achieved. The key to realizing such a computing environment is how to construct a shared memory system on the photonic ring. Thus, we propose and evaluate a shared memory system suitable to the virtual optical ring network, which takes into account contention resolution, cache coherency, and synchronization methods because the propagation delays among nodes are much larger than the conventional 1 shared memory system. Through simulation experiments of using three benchmark programs as representative parallel computing applications, we show the applicability of our shared memory system on the wide-area virtual photonic ring. Our simulation results show that the shared memory and access method for λ computing environment are effective in parallel computation for some of benchmark programs when the number of parallel nodes is less than eight and problem size is large. Furthermore, we propose and evaluate the method to improve the synchronization between computing nodes to enable collaboration. One of the obtained result is that the method can reduce 40% of the execution clocks when we use one of benchmark program.