This study seeks to understand how network failures affect the availability of service delivery across wide area networks and to evaluate classes of techniques for improving end-to-end service availability. Using several large-scale connectivity traces, we develop a model of network unavailability that includes key parameters such as failure location and failure duration. We then use trace-based simulation to evaluate several classes of techniques for coping with network unavailability. We find that caching alone is seldom effective at insulating services from failures but that the combination of mobile extension code and prefetching can improve average unavailability by as much as an order of magnitude for classes of service whose semantics support disconnected operation. We find that routing-based techniques may provide significant improvements, but that the improvements of many individual techniques are limited because they do not address all significant categories of network failures. By combining the techniques we examine, some systems may be able to reduce average unavailability by as much as one or two orders of magnitude.

Regarding finite state machines as Markov chains facilitates the application of probabilistic methods to very large logic synthesis and formal verification problems. In this paper we present symbolic algorithms to compute the steady-state probabilities for very large finite state machines (up to 10 27 states). These algorithms, based on Algebraic Decision Diagrams (ADDs) --- an extension of BDDs that allows arbitrary values to be associated with the terminal nodes of the diagrams --- determine the steady-state probabilities by regarding finite state machines as homogeneous, discrete-parameter Markov chains with finite state spaces, and by solving the corresponding Chapman-Kolmogorov equations. We first consider finite state machines with state graphs composed of a single terminal strongly connected component; for this type of systems we have implemented two solution techniques: One is based on the Gauss-Jacobi iteration, the other one is based on simple matrix multiplication. Then we...

Powerful workstations have become widely available as sources of computing cycles.

Obtaining performance models, like Markov chains and queueing networks, for systems of significant complexity and magnitude is a di#cult task that is usually tackled using human intelligence and experience. This holds in particular for performance models of a highly irregular nature. In this paper we argue by means of a non-trivial example --- a plain-old telephone system (POTS) --- that a stochastic extension of process algebra can diminish these problems by permitting an automatic generation of Markov chains. We introduce a stochastic process algebra that separates the advance of time and action occurrences. For the sake of specification convenience we incorporate an elapse operator that allows the modular description of time constraints where delays are described by continuous phase-type distributions. Using this language we provide a formal specification of the POTS and show how a stochastic process of more than 10 7 states is automatically obtained from this system description. ...

this paper may be copied or distributed royalty free without further permission by computer-based and other information-service systems. Permission to republish any other portion of this paper must be obtained from the Editor.

The performance of distributed text document retrieval systems is strongly in uenced by the organization of the inverted index. This paper compares the performance impact on query processing of various physical organizations for inverted lists. We present a new probabilistic model of the database and queries. Simulation experiments determine which variables most strongly influence response time and throughput. This leads to a set of design trade-offs over a wide range of hardware configurations and new parallel query processing strategies.

Regarding finite state machines as Markov chains facilitates the application of probabilistic methods to very large logic synthesis and formal verification problems. Recently, we have shown how symbolic algorithms based on Algebraic Decision Diagrams may be used to calculate the steadystate probabilities of finite state machines with more than 10 8 states. These algorithms treated machines with state graphs composed of a single terminal strongly connected component. In this paper we consider the most general case of systems which can be modeled as state machines with arbitrary transition structures. The proposed approach exploits structural information to decompose and simplify the state graph of the machine. 1 Introduction Finite state machines (FSMs), or their extensions, are often employed to model real digital systems for formal verification. As the complexity of those systems increases, probabilistic approaches to design and implementation verification become of interest; for...

The performance of distributed text document retrieval systems is strongly in uenced by the organization of the inverted index. This paper compares the performance impact on query processing of various physical organizations for inverted lists. We present a new probabilistic model of the database and queries. Simulation experiments determine those variables that most strongly in uence response time and throughput. This leads to a set of design trade-o s over a wide range of hardware con gurations and new parallel query processing strategies. Key Words. Performance, le organization, query processing, inverted le, inverted index, striping, shared-nothing, full text information retrieval. 1

Uniformization has been shown to be, in many cases, a good method to compute transient state probabilities of a continuous-time Markovchain. However, two issues limit its use: uniformization can be computationally very intensive, for instance, on sti# models, and uniformization cannot be used for all model classes, e.g., models with not uniformly bounded transition rates. In this paper weintroduce adaptive uniformization,avariation on standard uniformization, which can overcome these problems for some models. Adaptive uniformization di#ers from standard uniformization in that it uses a uniformization rate that adapts depending on the set of states that the process can be in after a particular number of jumps. Doing this can sometimes signi#cantly reduce the computational cost required to obtain a solution. A formal de#nition of adaptive uniformization is #rst given, along with a proof that adaptive uniformization yields correct results. Characteristics of models that can facilitate solution and alternative methods for computing the required #jump probabilities" are then discussed. Finally, the computational cost of adaptive uniformization #relative to standard uniformization# is illustrated, through its application to an extended machine-repairman model.

Redundancy techniques are commonly used to design dependable systems to ensure high reliability, availability and data integrity. Triple Modular Redundancy (TMR) is a widely used redundancy technique that masks faults. In a TMR system, we have three implementations of the same logic function and their outputs are voted using a voter circuit. In this paper, we present a new voter design called the Word-Voter that has some distinct advantages over the bit-by-bit voting schemes used in conventional TMR systems. This paper demonstrates the usefulness of the word-voter design in increasing the data integrity (reducing the probability of corrupt outputs) of TMR systems. The area and delay overhead of the word-voter design is compared to that of the bit-by-bit voter. An efficient design of a TMR-Simplex system using the word-voter is also presented.