/ Hidesada Kanda / Professor
/ Leo Nagamatsu / Associate Professor
/ Akira Fujitsu / Assistant Professor
/ Tongjun Huang / Research Associate
ISTC consists of the Information Processing Center and the University Library. ISTC's main purpose is to provide information services through computer facilities and library services. Our activities in 1995 are as follows.
Therefore, ISTC replaced about 400 workstations from the professors' offices, computer exercise rooms No. 1 and 2, LML and CAI rooms, the administarative offices, and each library lounge with causing minimum interruption. MUSE was revised to match the new systems.
Refereed Journal Papers
Leo Nagamatsu and Koichiro Deguchi. A performance evaluation model for multithreaded processor systems. Transactions of the Society of Instrument and Control Engineers, 32(6), 1996.
This paper proposes a performance estimation method for multithreaded processor systems. Multithreaded processors are known as effective memory latency tolerating scheme with multiple hardware contexts in a processor. It is important to estimate the processor working ratio for comparing various configurations. The proposed method is based on, firstly, assuming independent cache miss occurrences, and modeling the processor status with number of runnable threads and status of the thread in the execution stage, then employing a probability transition matrix between all possible processor statuses. Our method drastically reduced the amount of computation required to estimate the working ratio than traditional simulations.
It is very important for the effective use of highly parallel computer to understand its overall performance feature. But it depends on many elements, and shows very complicated behavior. We propose a new method to predict the overall performance using simple models built from a combination of the degree of parallelism, cache mechanism and capacity, communication cost and so on. Experimental results of performance evaluation on some real highly parallel computers show the feasibility of this method.
Leo Nagamatsu, Sanehiro Furuichi, and Koichiro Deguchi. Is double sized problem completed in double sized time? In Report of 37th Programming Symposium, Information Processing Society of Japan, 1996.
When the size of target problem is scaled up from small to large, how the required computation time will change? The behavior of achieved performance ( i.e. speed of computation ) is rather complex, since it is result of interaction between required and available resource sizes. In study of performance evaluation of parallel computers, this behavior is simplifyed and modeled using a parameter called SCALABILITY. In this paper, we propose a new method for modeling relation between size of computation and expected performance, and give generalized extension of scalability. This method can provide guidelines for effective utilization of parallel computer systems on various conditions.
In this paper, we present a new communication structure called k-layered coterie which is suitable for the multi-round communication. Based on the k-layered coterie, we propose a new distributed algorithm with $dNlog_d(N)$ message passes and $log_d(N)$ rounds of communication, where $d$ is an integer for decentralized consensus problem.
In this paper, we propose a new distributed algorithm for implementing first-order multiparty interactions based on a new implementation model, in which a role manager is devised for each role. Our algorithm is closer to the fully distributed one than that in Joung's algorithm in the sense that (1) every process (role manager) only knows its adjacent role managers (processes), and (2) the time complexity of local computation in each process and role manager is equal.