Ravi Jain

Mobile Internet access and QoS guarantees using mobile IP and RSVP with location registers

The mobile IP (MIP) protocol for IP version 4 as being standardized by the Internet community provides continuous Internet connectivity to mobile hosts, without requiring any changes to existing routers and higher-layer applications. We propose an alternative protocol, mobile IP with location registers (MIP-LR) which is closer to the service node database approach used in the public switched telephone network (PSTN): before launching a packet to the mobile host, the sender first queries a database to obtain the recipients current location. MIP-LR is designed for operation in enterprise environments or within logical administrative domains, as it requires a sending host to be aware of which hosts are potentially mobile and implement the MIP-LR protocol. The benefits of MIP-LR are that potentially long routes, called triangle routes, from the sender to the mobile host are avoided, encapsulation of packets sent to a mobile host is not required, the load on the home network as well as the home and foreign agents is reduced, and there is substantially improved interoperability with protocols such as RSVP for providing QoS guarantees. We carry out a simplified average-case analysis of the costs and benefits of MIP-LR and show it can result in significant reductions in mean network costs compared to MIP.

Heuristics for scheduling I/O operations

The I/O bottleneck in parallel computer systems has recently begun receiving increasing interest. Most attention has focused on improving the performance of I/O devices using fairly low level parallelism in techniques such as disk striping and interleaving. Widely applicable solutions, however, will require an integrated approach which addresses the problem at multiple system levels, including applications, systems software, and architecture. We propose that within the context of such an integrated approach, scheduling parallel I/O operations will become increasingly attractive and can potentially provide substantial performance benefits. We describe a simple I/O scheduling problem and present approximate algorithms for its solution. The costs of using these algorithms in terms of execution time, and the benefits in terms of reduced time to complete a batch of I/O operations, are compared with the situations in which no scheduling is used, and in which an optimal scheduling algorithm is used. The comparison is performed both theoretically and experimentally. We have found that, in exchange for a small execution time overhead, the approximate scheduling algorithms can provide substantial improvements in I/O completion times.

Scheduling parallel I/O operations in multiple bus systems

Abstract The I/O subsystem was one of the first areas of computer system design to incorporate parallelism. However, enhancement of the parallelism of I/O systems has received little attention in parallel system design. There has been almost no study of the benefit of scheduling parallel I/O operations to increase the multiprocessor system performance. An I/O transfer typically requires a processor or memory, a channel, and an external memory device simultaneously. Parallel I/O thus requires scheduling multiple resources simultaneously, rather than a single resource serially. This paper presents an algorithm for optimal scheduling of batched parallel I/O requests for a common class of shared memory multiprocessors. The algorithm is essentially an optimal k-coloring of a bipartite graph with arbitrary edge weights, where the vertices represent processors and memories and the edges represent I/O transfers. The best previously known k-coloring algorithm has running time O(n5). We show a series of improvements to obtain an algorithm with a running time O(n3(log n + log K)), where n is the number of vertices and K is the maximum edge weight, i.e., the length of the longest I/O transfer. We conclude with an experimental study of the performance of the algorithms.

Scheduling on airdisks: efficient access to personalized information services via periodic wireless data broadcast

There has been considerable interest in delivering information to distributed mobile clients via wireless broadcast. Information transmitted periodically over wireless media can be regarded as a virtual disk, which we call an airdisk, analogous to a standard magnetic disk. Airdisks offer an efficient mechanism for delivering personalized information services to mobile clients with portable or laptop computers by broadcasting data and allowing clients to filter out the items of interest to them. We study the problem of scheduling the order in which data items are broadcast so as to minimize the access time of the clients, focusing on the case where the server inserts an index at the start of the broadcast period. We observe that the problem is analogous to that of determining how data should be laid out on the disk, and show that the problem is in general NP-complete. We develop a branch-and-bound procedure for solving the problem optimally, and then develop a fast, simple heuristic. The results of our simulation experiments show that the heuristic runs substantially faster than the branch-and-bound procedure, and yet produces schedules that are only slightly longer.

Input/Output in Parallel and Distributed Computer Systems

Input/Output in Parallel and Distributed Computer Systems has attracted increasing attention over the last few years, as it has become apparent that input/output performance, rather than CPU performance, may be the key limiting factor in the performance of future systems. This I/O bottleneck is caused by the increasing speed mismatch between processing units and storage devices, the use of multiple processors operating simultaneously in parallel and distributed systems, and by the increasing I/O demands of new classes of applications, like multimedia. It is also important to note that, to varying degrees, the I/O bottleneck exists at multiple levels of the memory hierarchy. All indications are that the I/O bottleneck will be with us for some time to come, and is likely to increase in importance. Input/Output in Parallel and Distributed Computer Systems is based on papers presented at the 1994 and 1995 IOPADS workshops held in conjunction with the International Parallel Processing Symposium. This book is divided into three parts. Part I, the Introduction, contains four invited chapters which provide a tutorial survey of I/O issues in parallel and distributed systems. The chapters in Parts II and III contain selected research papers from the 1994 and 1995 IOPADS workshops; many of these papers have been substantially revised and updated for inclusion in this volume. Part II collects the papers from both years which deal with various aspects of system software, and Part III addresses architectural issues. Input/Output in Parallel and Distributed Computer Systems is suitable as a secondary text for graduate level courses in computer architecture, software engineering, and multimedia systems, and as a reference for researchers and practitioners in industry.

Applying randomized edge coloring algorithms to distributed communication: an experimental study

We propose a parameterized, randomized edge coloring algorithm for use in coordinating data transfers in fully connected distributed architectures such as parallel 1/0 subsystems and multimedia information systems. Our approach is to preschedule 1/0 requests to eliminate contention for 1/0 ports while maintaining an efficient use of bandwidth. Request scheduling is equivalent to edge coloring a bipartite graph representing pending 1/0 requests. Although efficient optimal algorithms exist for centralized edge coloring where the global request graph is known, in distributed architectures heuristics must be used. We propose such heuristics and use experimental analysis to determine their ability to approach the centralized optimal. The performance of our algorithms is also compared with the work of other researchers experimentally. Our results show that our algorithms produce schedules within 5% of the optimal schedule, a substantial improvement over existing algorithms. The use of experimental analysis allows us to evaluate the appropriateness of each heuristic for a variety of different architectural models and applications.

Distributed scheduling algorithms to improve the performance of parallel data transfers

The cost of data transfers, and in particular of I/O operations, is a growing problem in parallel computing. A promising approach to alleviating this bottleneck is to schedule parallel I/O operations explicitly. We develop a class of decentralized algorithms for scheduling parallel I/O operations, where the objective is to reduce the time required to complete a given set of transfers. These algorithms, based on edge-coloring and matching of bipartite graphs, rely upon simple heuristics to obtain shorter schedules. We present simulation results indicating that the best of our algorithms can produce schedules whose length is within 2--20% of the optimal schedule, a substantial improvement on previous decentralized algorithms. We discuss theoretical and experimental work in progress and possible extensions.

Analysis of approximate algorithms for edge-coloring bipartite graphs

The problem of edge-coloring a bipartite graph is to color the edges so that adjacent edges receive different colors. An optimal algorithm uses the minimum number of colors to color the edges. We consider several approximation algorithms for edge-coloring bipartite graphs and show tight bounds on the number of colors they use in the worst case. We also present results on the constrained edge-coloring problem where each color may be used to color at most k edges.

A GRAPH-THEORETIC MODEL FOR THE SCHEDULING PROBLEM AND ITS APPLICATION TO SIMULTANEOUS RESOURCE SCHEDULING

ABSTRACT As computer and communications systems become more complex, system designers are broadening their concern from the scheduling of individual resources to scheduling tasks which require multiple resources simultaneously. However there are relatively few published results for simultaneous resource scheduling problems. Equally importantly, although there are algorithms for special instances of these problems in different application areas, there is no general model that provides a framework for applying results across application boundaries. In this paper we present a graph-theoretic model for formally specifying scheduling problems. We apply the model to several simultaneous resource scheduling problems, drawn from the increasingly important application area of scheduling input/ouput (I/O) tasks in a parallel computer system. The model is thus used to specify the I/O scheduling problem for parallel computers with tree architectures, parallel I/O in the presence of a class of mutual exclusion constraints, and scheduling parallel I/O for computers with multiple I/O bus architectures. We use the model for recognition of the similarity between scheduling problems, the transformation of one scheduling problem to a different scheduling problem, and the decomposition of a problem into subproblems. We thus obtain an optimal algorithm for parallel I/O in tree architectures, an optimal algorithm for scheduling parallel I/O in the presence of limited mutual exclusion constraints, and an improved heuristic for parallel I/O in multiple bus architectures.

Scheduling parallel I/O operations

The I/O bottleneck in parallel computer systems has recently begun receiving increasing interest. Most attention has focused on improving the performance of I/O devices using fairly low-level parallelism in techniques such as disk striping and interleaving. Widely applicable solutions, however, will require an integrated approach which addresses the problem at multiple system levels, including applications, systems software, and architecture. We propose that within the context of such an integrated approach, scheduling parallel I/O operations will become increasingly attractive and can potentially provide substantial performance benefits.We describe a simple I/O scheduling problem and present approximate algorithms for its solution. The costs of using these algorithms in terms of execution time, and the benefits in terms of reduced time to complete a batch of I/O operations, are compared with the situations in which no scheduling is used, and in which an optimal scheduling algorithm is used. The comparison is performed both theoretically and experimentally. We have found that, in exchange for a small execution time overhead, the approximate scheduling algorithms can provide substantial improvements in I/O completion times.