Gita Gopal

The Case for Orderly Sharing

Transaction processing systems with the functionality and throughput desirable to support future telecommunications network applications appear to be beyond the capability of conventional multiprocessor architectures. A central obstacle to realizing such systems is the cost of coordinating access of transactions in the various processors to shared data. In this paper, we introduce the principle of orderly sharing — the efficient coordination of many loosely coupled processing elements acting on shared resources (e.g., data items) through orderly distribution of information regarding the state of resources throughout the system. We illustrate the application of this principle in the context of a novel system called the Datacycle architecture [Herm87a].

Achieving throughput and functionality in a common architecture: the Datacycle experiment

Summary form only given. The Datacycle database architecture is based on repetitive broadcast of database contents over high bandwidth channels and on-the-fly filtering of database contents to extract records of interest. These two mechanisms-high speed broadcast and on-the-fly filtering-result in a system that achieves the goals of high performance transaction processing, a powerful and flexible query capability, and high levels of concurrent access to a single database by multiple applications. The prototype system comprises multiple pump boards, each of which stores up to 128 Mbytes of database contents, and multiple access manager boards, each of which contains three on-board VLSI data filter chips. Each VLSI filter executes queries against the contents of a 53 Mbyte/second broadcast channel; selected records and/or computed aggregate values are returned to the host application.<<ETX>>

Limited waiting: an adaptive overload-control strategy for circuit switched networks

The authors present an adaptive control strategy for circuit-switched networks in which calls are permitted to wait for a limited time to be carried on a low-cost preferred route. The strategy uses network resources more efficiently than other conventional approaches, exhibiting lower blocking of calls under overload conditions. The proposed strategy is also adaptive in that, when there is little network congestion, there is no penalty for the policy, in contrast to other schemes in which improved performance under overload is achieved at the cost of degraded performance at lower bounds. The authors analyze an exactly solvable model and present simulation results for a more general network. >

Software fault tolerance in telecommunications systems

1. I n t r o d u c t i o n The telephone system in many countries including the United States depends on stored program control switching systems to establish calls between millions of users. The programs for these systems are large and expensive to develop and maintain. The size and complexity of the programs arise from various requirements [1]: • high functionality • strict real-time constraints (30 msec response time to a request) • high traffic levels (100-200 call attempts/see) • a variety of hardware devices to be controlled • very high availability (down 2 hours in 40 years) and reliability (2 lost calls in 10,000) requirements • ability to run unattended for long periods. Complexity will increase rapidly with the network trend toward a more standard and open software environment, supporting: • more complex interactions among network elements such as switching nodes which perform call processing, signaling transfer points which allow switching nodes to communicate, and service control points which control data accessed by switching nodes • more equipment suppliers, leading to greater system heterogeneity • greater complexity within switching nodes, due to open interfaces and multiple application developers within a single system (instead of just one or two as in the past) • highly distributed intelligence and rapid introduction of new features and services. Reliability has been a major issue in tile design of telecommunications systems. Hardware fault tolerance is well understood, and arbitrarily small hardware error probabilities can be obtained by combining techniques such as modular fail-stop design, redundancy, automatic reconfiguration and

Routing in a circuit-switched network with priority classes

The problem of routing in a circuit-switched network with different classes of traffic is examined. The authors consider the case of two priority classes, assuming that a relatively small fraction of traffic comprises the high-priority class requiring a better grade of service, with the rest being treated as lower priority normal traffic. These high-priority calls might be used to provide a premium service with lower blocking probability, or to implement services in which failure is expensive to the network such as a multiparty conference call. The authors identify several routing schemes that give preference to high-priority calls, either by explicitly reserving capacity or by allowing more resource sharing. The performance of these schemes is studied in terms of blocking probabilities for the two classes, protection of priority traffic against overload of normal traffic, and adaptability to variations in offered priority traffic. The authors compare the schemes by the different total revenues they earn and discuss the implementation complexity. They find that some resource-sharing schemes display good performance, having the potential to provide low blocking for priority calls with minimal impact on the remaining normal traffic.<<ETX>>

Distributed implementation of real-time resource counters

The problem of managing telephone network resources for private virtual networks is discussed. Implementing a real-time resource counter on a distributed system is investigated. A correctness condition is developed for a real-time resource counter, and several approaches to implementing it are described. The various approaches are evaluated on a model that accounts for load-dependent message-processing delays, using both analytic and simulation techniques. The results of the performance studies also suggest heuristics for configuring a resource counter on a distributed system. A very simple approach, in which requests that cannot be satisfied locally are randomly forwarded to other processors, works well over a wide range of loads and system sizes. Unless a very large number of processors is required to handle the requests for resources or heavy overloads are expected, this may be the best algorithm.<<ETX>>