LIBRARY OF THE UNIVERSITY OF ILLINOIS AT URBANA-CHAMPAIGN 5/0.84 „o. (o0\-t>0 uiucDcs-R-73-606 7x^^ A SUMMARY OF CURRENT RESEARCH IN COMPUTER NETS November, 1973 by Edward K. Bowdon, Sr. DEPARTMENT OF COMPUTER SCIENCE UNIVERSITY OF ILLINOIS AT URBANA-CHAMPAIGN URBANA, ILLINOIS THE LIBRARY OF THE UNIVERSITY OF ILLINOIS uiucdcs-r-73-6o6 A SUMMARY OF CURRENT RESEARCH IN COMPUTER NETS by Edward K. Bowdon, Sr. November, 1973 Department of Computer Science University of Illinois at Urbana-Champaign Urbana, Illinois This research was supported in part by the National Science Foundation under Grant No. GJ 28289. 1. INTRODUCTION From a postulation of the essential characteristics of a network computer, we have developed a queueing theory model for a multiserver system with a finite length priority queue [5]. Assuming Poisson input and exponentially distributed processing times, we have utilized the average waiting time results from the model to determine priority assignment and steady state job dispatching rules for the network [3,^,10]. We have developed a GPSS model to evaluate current computer center operations at the University of Illinois [9]. The multi-faceted priority scheme presently used to schedule jobs on the IBM 360/75 has been evaluated and we have proposed a new priority scheme [7]. The new scheme assigns an initial static priority to a job based on its total O.S. time requirement. There are basically two priority classes determined by this initial assignment and jobs are ranked within their respective classes. This initial assignment is dynamically adjusted depending on l) the ratio of l/0-bound to CPU-bound jobs in the system and 2) total system balance as measured by an evaluation of the processor and memory usage. We have developed a measure of cost effectiveness for evaluating computer center throughput [2]. We have determined that our measure of cost effectiveness is very indicative of the level of activity in computing centers and the turnaround time for jobs in different priority classes. We have formulated an algorithm which allows the user to specify a series of deadlines for a job and associate with each deadline a relative reward. We have formulated, for each center, heuristic scheduling and priority assignment algorithms which guarantee that the deadlines for the highest priority jobs will be met, where possible [6]. The algorithm further indicates those deadlines which are in danger of being missed. We then use the measure of cost effectiveness to perform load-leveling between centers and, as a result, achieve economic viability within the network computer. 2. COMPUTER NETWORK MODELING We began this research effort by examining several existing networks. These networks were selected for study because they represent the wide variety of philosophies and ideas which enter into the design of network computers. The problems of interest here include the relative capabilities of different network configurations, identification of specific limitations of different networks, and the interrelationship between communication and computing. From a long range viewpoint, one of the more interestiag problems is the effect on system performance of centralized vs. distributed control in the operating systems. We are developing queueing theory models for several of the existing and proposed network computers (Aloha, AREA, DCS, and TUCC). Our view of a typical network is much the same as an outside observer's. We see jobs arrive at the i center with rate a. and eventually leave, either through process completion at rate \±. or through transmission to another center (for whatever reason) at rate v.. Internal to this structure is the 1 priority assignment, if any, and the load leveler for that particular center. Each center is considered to be of equal, although not necessarily homogeneous, capacity. Hyperexponential arrival and service rates are being considered as well as Poisson arrivals and exponential service times. Mills has postulated a queueing theory model of a theoretical network and used this model to predict such measures as the number in the system, the number in the queue, and system efficiency [12]. Under the assumptions of Poisson arrivals, exponential service times and a restricted k queue length, the mathematical model is utilized to develop the steady- state equations for the two-center and the n-center network. Then, a priority scheme is implemented and the steady state equations are rederived. With these two varying assumptions, conclusions are drawn as to their effect on system efficiency and the number in the queue. Lastly, the relevance of proposing an analytical model is discussed. The mathematical model is compared to existing networks to verify its authenticity and accuracy. The theory of queueing systems at steady state is more advanced than that of systems at arbitrary points in time for two reasons. First, many systems are assumed to approximate steady state after a relatively short period of time. Second, steady state analysis is easier to do than time -dependent analysis. While steady state analysis provides much useful information about queueing systems, it gives no indication of how long it takes a system to ready steady state or how the system behaves before reaching it. Further, many systems never reach steady state, so steady state analysis cannot be applied at all in these cases. Barr has defined time -dependent system descriptors which are analogous to some well-known steady state descriptors for a Poisson system with infinite queue length [1]. The transient solutions for these systems are used to determine the Laplace transforms of the transient descriptors explicitly by inverting their Laplace transforms numerically. These descriptors are then used to determine the effect on the network when one or more computers goes down temporarily and then returns to service. 3. CENTER THROUGHPUT ANALYSIS We are investigating dynamic priority assignment rules for efficient job processing, including those algorithms which minimize the mean flow time while maintaining system balance. We have developed a GPSS model for ILLINET (the computer communication network at the University of Illinois). Using actual data from the network we have shown that a proposed dynamic priority assignment algorithm yields better throughput than the existing algorithm and at the same time maintains a higher level of resource utilization. We have also used the network simulator to demonstrate the efficacy of networking and the merits of load leveling between centers. The present scheduling algorithm queues job for service according to their resource requests [7]. Very important tasks must assume queue positions determined by their particular set of resource requests, irrespective of their urgency. A priority algorithm is desired, therefore, that allows jobs to vie for queue position depending on how short a turnaround time the user desires for his job. The basic principle behind such an algorithm is that users would be able to choose a priority for their jobs and would be charged for service accordingly. If a user is willing to pay more than the normal rate his job would be given a higher priority; if a user wishes to pay less than the normal rate, his job would be given a lower priority. This sort of scheduling algorithm is referred to as "pay-for-priority. " A GPSS simulation of the IBM 360/75 was used to implement and evaluate several specific pay-for-priority schemes [11]. In order to use the GPSS simulator to predict real system performance within well defined statistical limits of accuracy, the simulator had to be tuned with data characteristic of the environment to which the results were to be applied. Therefore, a meaningful benchmark job stream representing the workload of the period under study had to be created. Formulating a tuned simulator from the various system performance parameters produced by running the benchmark, and then validating the accuracy of the simulator were preliminary to any actual pay-for-priority algorithm evaluations. As the final step in the pay-for-priority scheme development and evaluation, simulation runs of eight specific models were performed and analyzed, yielding comparative performance data. k. SYSTEM PERFORMANCE EVALUATION The growing number of experimental network computers has precipitated the need for some method of evaluating these networks. In order to establish requirements for advanced computer communications networks we are determining measures of cost effectiveness to facilitate studying the effects of network size on system performance and evaluating the utility of increased complexity. We have begun this effort by using simulation techniques to evaluate the adequacy of using throughput, turnaround time, and resource utilization in measuring system performance in our models. Until recently, efforts to measure computer system performance have centered on the measurement of resource (including processor) idle time. A major problem with this philosophy is that it assumes that all tasks are of roughly equal value to the user and, hence, to the operation of the system. Our research in this area has been aimed at evaluating the efficacy of priority assignment and job dispatching rules for networks consisting of IBM 360/75* s run under HASP 3-1 and O.S. We have developed a simulation model for a hypothetical geographically distributed network computer [7]. Since the model was developed for a hypothetical network, we needed to ensure that the results were valid and That no gross errors existed in the model. Our approach was to design a general n node network simulator and then to particularize the input parameters to describe ILLINET. For a given period, system accounting records provided exact details of the resources used by each task in the system including CPU usage, input/output resources used, core region size requested, and total real time in the system. Using the first three of these parameters as input 8 data, we could simulate the fourth. Comparison of the actual real time in the system to the simulated real time in the system authenticated the accuracy of the model. Extrapolating from these results, we could then consider the more general network with reasonable assurance of accurate results. We have obtained some interesting results from using simulation as a tool for evaluating alternative methods of improving system performance in network computers [8]. While using actual data from the network to verify that the model accurately predicts each job's total time in the system, we evaluated the resource utilization. Then we used the simulator to show that a proposed dynamic priority assignment algorithm yields better throughput than the existing algorithm while maintaining a higher level of CPU and memory utilization. As computing needs increase, the resources at a given installation become inadequate to satisfy the requirements of all users of the facility. A network computer offers virtually an unlimited capability for expanding computer resources without the costs involved in having all of the equipment at the individual installation. Since the needs of computing centers vary with time, idle resources normally existing during a light load period can be utilized by another overloaded center. Salz has investigated the benefits obtainable through the use of networking [13 ]• A "pay- for -priority" scheduling algorithm for the individual centers is explored and load leveling techniques for transmitting jobs between centers are discussed. Finally, the effects of these algorithms on system performance are demonstrated through the use of a simulation model. 5. REFERENCES [1] Barr, R. H., "Time -Dependent Descriptors for the Poisson Queue," Master of Science Thesis, Department of Computer Science, University of Illinois at Urbana -Champaign, Urbana, Illinois, Report No. UIUCDCS-R-73-579, June, 1973. [2] Barr, W. A., "Cost Effective Analysis of Network Computers," Master of Science Thesis, Department of Computer Science, University of Illinois at Urbana -Champaign, Urbana, Illinois, Report No. UIUCDCS-R-72-538, August, 1972. [3] Bowdon, E. K., Sr., "Priority Assignment in a Network of Computers," IEEE Transactions on Electronic Computer, November, 1969. [k] , "Dispatching in Network Computers," Proceedings of the International Symposium on Computer-Communications Networks and Teletraffic, April, 1972. [5] , "Network Computer Modeling," Proceedings of the ACM Conference, August, 1972. [6] and W. J. Barr, "Cost Effective Priority Assignment in Network Computers," Proceedings of the FJCC, December, 1972. [7] , S. A. Mamrak and F. R. Salz, "Simulation: A Tool for Performance Evaluation in Network Computers, " in Proceedings of the N.C.C. & E, June, 1973 . [8] , S. A. Mamrak and F. R. Salz, "Performance Evaluation in Network Computers, " in Symposium on the Simulation of Computer Systems, June 19-20, 1973. [9] , S. A. Mamrak and F. R. Salz, "A Simulation Tool for Performance Evaluation of the IBM 360/75, " International Journal of Computer and Information Sciences, Vol. 3, No. 1, 197^ (in press). [10] Fitzgerald, J. T., "Load Regulation and Dispatching in a Network of Computers, " Master of Science Thesis, Department of Computer Science, University of Illinois at Urbana -Champaign, Urbana, Illinois, Report No. UIUCDCS-R-72-537, August, 1972. [11] Mamrak, S. A., "Simulation Analysis of a Pay-for-priority Scheme for the IBM 360/75," Master of Science Thesis, Department of Computer Science, University of Illinois at Urbana-Champaign, Urbana, Illinois, Report No. UIUCDCS-R-73-605, August, 1973. 10 [12] Mills, L. A., "Queues and Network Computers," Master of Science Thesis, Department of Computer Science, University of Illinois at Urb ana -Champaign, Urbana, Illinois, Report No. UIUCDCS-R-73-577, May, 1973. [13] Salz, F. R., "Simulation Analysis of a Network Computer," Master of Science Thesis, Department of Computer Science, University of Illinois at Urbana-Champaign, Urbana, Illinois, Report No. UIUCDCS-R-73-578, May, 1973. BLIOGRAPHIC DATA EET 1. Report No. UIUCDCS-R-73-606 3. Recipient's Accession No. Title and Subtitle A SUMMARY OF CURRENT RESEARCH IN COMPUTER NETS 5- Report Date November, 1973 Author(s) Edward K. Bowdon, Sr. 8. Performing Organization Rept. No. Performing Organization Name and Address Department of Computer Science University of Illinois Urbana, Illinois 6l801 10. Project/Task/Work Unit No. 11. Contract/Grant No. NSF C-J 28289 Sponsoring Organization Name and Address National Science Foundation Washington, D.C. 13. Type of Report & Period Covered Research u. . Supplementary Notes Presented at the 7th Hawaii International Conference on Systems Sciences. . Abstracts In this paper we present some recent developments in the research program being conducted at the University of Illinois to formulate analytical tools for system modeling and analysis of real time computer networks. Specifically, we are conducting theoretical studies of geographically distributed network computers in three separate and distinct areas: Computer Network Modeling, Center Throughput Analysis, and System Performance Evaluation. Taken together these three areas provide the computer systems analyst with the necessary tools for modeling, analyzing, and evaluating system performance in existing and proposed network configurations - a first step in the long road to achieving economic viability in network computers. Key Words and Document Analysis. 17a. Descriptors Computer Networks Modeling and Analysis Center Throughput Performance Evaluation b. Idcntif icrs /Open-F.ndcd Terms c. ' OSATI Fie Id/Group ■•Availability Si atement Release Unlimited 19. Security Class (This Report) UNCLASSIFIED 20. Security Class (This Page UNCLASSIFIED 21. No. of Pages 10 22. Pric. |RM M TIS-39 (10-70) USCOMM-DC 40329-P7I # s* * <0 UNIVERSITY OF ILLINOI9-URBANA 3 0112 047417826