Hikaru Nakanishi

Adaptive load balancing with bidirectional piggybacking

Computer network technology is no longer experimental. Having come of age, this technology is now both practical and necessary. Downsizing of computers and open and high-speed networks have required multimedia processing as well as traditional data processing on a networked system connected by various types of computers. However, in current network systems, users are responsible for job submission, while simultaneously considering what and where the necessary and available resources are. This will create a heavier burden in the future. A framework of load balancing with distributed dispatchers that uses an analysis of network delay and job characteristics is proposed here. Examining an adaptive load balancing scheme with bidirectional piggybacking permits basic study of this framework. First a bidirectional piggybacking is proposed that uses the results of a job as an information carrier for a remote server and that uses the request of a job as an information carrier for a local server. We also propose a load balancing scheme using bidirectional piggybacking. This scheme is then compared with other schemes on homogeneous and heterogeneous environments with and without communication delay. Finally, the reliability of piggy-backed information on the arrival rate of local/remote job, communication delay, etc. are discussed. The effects of this scheme based on bidirectional piggybacking are also discussed.

Approximate performance evaluation of parallel processing systems using average concurrency

For the performance evaluation of the parallel processing system of a set of concurrent programs (program group), several methods have been proposed such as decomposition approximation, which consider the program group with fixed concurrency. On the other hand, this paper proposes an innovative approximation method considering a more general program group in which the concurrency changes for each program or with time. First, the average concurrency is determined which is the average of the concurrency of the program group. Then the decomposition approximation is applied to two integers giving bounds to the average concurrency, and the characteristic parameters such as throughput are determined. The two obtained sets of characteristic parameters are averaged with weight, providing the evaluation for the given system. By this method, a system which cannot be handled by the decomposition approximation can now be evaluated. The method has an advantage that the computation can be performed with a very small number of state spaces. Because of this property, the parallel processing system can be evaluated with a high speed and a small memory. Finally, the results obtained by the proposed method and the results of discrete event simulation technique by a supercomputer are compared for several models, and it is indicated that a sufficient accuracy is realized.

Performance of window flow control scheme for interconnected packet networks

When communication packet networks are interconnected, all internetwork traffic concentrates at the connecting point, and flow control is needed. X.75 provides window flow control at the connecting point (signaling terminal equipment). The authors investigate the performance of the window-flow-control scheme by simulation. They compare three ways of instituting flow-control-boundary division, and show that end-to-end within each network and between the two terminal is the best boundary division because its offers has satisfactory total delay with small buffer in the signaling terminal equipment. The concept and design of a service access procedure are presented for establishing transparent and reliable connections between the top three layer protocols of the ISO Open System Interconnection (OSI) and the corresponding bottom four layer protocols. The transport and network protocols have been selected to be the US Department of Defense standard transmission control protocol (IP), respectively. The upper layer protocols are made a part of the host processors software whereas the TCP/IP and the subsequent lower layer protocols are resident in a physically distinct device such as an interface unit.<<ETX>>

Permanent time‐stamp ordering mechanism

This paper proposes a permanent time-stamp ordering mechanism as a concurrency control metabolism in the distributed database, where transactions are served in the order of their arrivals. In the proposed mechanism in the read time stamp and the write time stamp are recorded on each data item; and when a conflict arises between transaction services, the transaction with a larger time stamp is aborted. This paper describes the algorithm for the permanent time-stamp ordering mechanism and confirms the validity of the algorithm

Modeling and analyzing advance processing control by extended message-passing

The job for which processing is determined only in the course of execution is called the unschedulable job. When an unschedulable job is to be executed in a parallel processing system, it is not predetermined when the communication between processors arises, and synchronization among processors is a serious problem. This paper considers the mathematical theory for the existence of interprocessor communication and the transmission time in the execution of the unschedulable job. The message-passing scheme is extended, considering the message transmission probability and the indeterminate aspects of the transmission time. Then the advance processing control is considered, which helps the synchronization among processors in the execution of the unschedulable job, and the overhead is theoretically analyzed. The condition for the task assignment to the processors is derived so that the deterioration of the processing speed due to the overhead can be minimized.

Approximate analytic method for computer systems with multiple level concurrent programs

Closed queuing network models representing computer systems with programs that have variation in concurrency level are discussed. For such models computation by existing approximation methods results in an explosion in the size of the state transition matrix. An innovative approximate analytic method based on the overall average concurrent level of programs is proposed. It is found that use of the proposed method considerably reduces the number of states and thus requires less memory. This average concurrency method is a fast algorithm for predicting performance levels of computer systems. Various examples were studied and the accuracy of the method was confirmed determined by comparison to the results of more exact simulations.<<ETX>>

Simulation language based on logic for queueing network using time object

Queueing network simulation (QNS) has recently been considered as one of the important means for performance evaluation of the network-structured system. Presently, FORTRAN, simulation language GPSS and other languages are used to describe QNS softwares, but their describability is not very high. This paper notes that the simulation condition defining the QNS model can be described primarily by conditional form, and the logical language PROLOG has a high describability. A QNS-dedicated logical language SILQ (Simulation Language based on Logic for Queueing network), is proposed where QNS model is defined by Prolog form. The language has a feature that the program can be constructed without considering a Prolog algorithm, while it has the disadvantage that the time-scale, which is indispensable in a simulation, cannot be represented. As a means to introduce the time-scale without degrading the describability, a time object is proposed which provides a parameter representation for the state transition in the system. Then the programming primitives are prepared to describe the operation and other events for each time object in Prolog form. Programming examples by SILQ are presented, and the result is compared with other languages, indicating the high describability of the proposed language. A sequential processing system is constructed for SILQ, and a parallel processing system is discussed, aiming at the improvement of the processing speed.

D‐SSQ: A queueing network simulator with parallel execution control

This paper discusses a node-distributed parallel-processing event-driven queueing network simulator, called D-SSQ. It employs the advanced execution control, aiming at the effective utilization of the parallelism inherent in the queueing network. The advanced control works as follows. Each processor first works ignoring the restrictions from other processors. If a contradiction is detected by manifesting the effects of other processors through the interprocessor communication, the contradiction is dissolved by partially cancelling the results of the previous processing. The processing performance of D-SSQ was examined by an experimental system and by the simulation using a large-scale computer. It is observed that the processing capability increases linearly with the number of processors, but does not increase as rapidly as was expected because of the large overhead in the distribution and the ineffective processing by the excessive advance control. Consequently, a restricted advance control has been considered, aiming at the elimination of the excessive advanced operation. It is indicated that the processing performance can be improved and there exists an optimum value for the extent of advanced operation.

Performance evaluation of asynchronous distributed processing system with limited‐advance processing

The traditional parallel processing assumes synchronous jobs, for which the processing can easily be planned before execution. On the other hand, the parallel processing is not usually considered for asynchronous jobs of a stochastic nature, such as the queueing network simulation, since the synchronization control of the processors is a problem. The advance processing has already been proposed as a control scheme for the processors in the asynchronous job parallel processing system. In this scheme, each processor performs the processing, thereby ignoring its relation to other processors. Thus the result of the processing is cancelled when a contradiction is found. Unfortunately, the processor may be too advanced in the execution, which results in a large number of cancelling, thus decreasing the processing ability per processor. From such a viewpoint, this paper discusses the limited-advance processing which is the improved scheme of advance processing. In this scheme, if the processor performs a certain amount of processing, it judges itself as too advanced, and awaits the progress of other processors. It is shown that there exists the optimum advance limit in this scheme, which is a function of the number of processors. The proposed scheme is analyzed to predict the optimum advance limit. In the analysis, the relation between the simulation time and the real time is specified, and the distribution overhead, cancelling time and the number of iterations are discussed. Comparing the result of numerical computation in the analysis and the result of measurement in the prototype system, the validity of the analysis is verified.

Delay Analysis of Wait System Model for Window-Controlled Packet Network and Optimal Window Allocation

Window allocation is an important problem in designing a window-controlled packet network. Most of the early studies on this topic deal with the loss system model because of its analytical simplicity. In actual packet switched networks the admission delay of an input-regulated packet has to be considered and hence its wait system model need be studied. This paper presents delay analysis of the wait system model for a multilogic channel network (multi-LC model) with each logic channel having its own window. First, the equivalent flow method is validated in a single-LC model. Then, based on this method, an approximate approach applicable to the multi-LC model is developed. Our approximate method is evaluated in comparison with simulation results. Then, using this approximate method, the optimal window allocation problem is studied in a general wait system packet network. We obtain an interesting result showing that the optimal allocation with minimum delay can be achieved by equalizing the input regulation degree on each LC.