Pankaj Mehra

Synthetic workload generation for load-balancing experiments

The Dynamic Workload Generator accurately replays measured workloads in the presence of competing foreground tasks. We have used this workload-generation tool to predict the relative speedups of different sites for an incoming task in our prototype system, using only the resource-utilization patterns observed before the task arrives. Our results show that the load-balancing policies learned by our system effectively exploit idle resources of a distributed computer system.Dynamic Workload Generator is a facility for generating realistic and reproducible synthetic workloads for use in load-balancing experiments. For such experiments, the generated workload must not only mimic the highly dynamic resource-utilization patterns found on todays distributed systems but also behave as a real workload does when test jobs run concurrently with it. The latter requirement is important in testing alternative load-balancing strategies, a process that requires running the same job multiple times, each time at a different site but under an identical network-wide workload.Parts of DWG are implemented inside the operating-system kernel and have complete control over the utilization levels of four key resources: CPU, memory, disk, and network. Besides accurately replaying network-wide load patterns recorded earlier, DWG gives up a fraction of its resources each time a new job arrives and reclaims these resources upon job completion. Pattern-doctoring rules implemented in DWG control the latter operation. This article presents DWGs architecture, its doctoring rules, systematic methods for adjusting and evaluating doctoring rules, and experimental results on a network of Sun workstations.

Automated Learning of Workload Measures for Load Balancing on a Distributed System

Load-balancing systems use workload indices to dynamically schedule jobs. We present a novel method of automatically learning such indices. Our approach uses comparator neural networks, one per site, which learn to predict the relative speedup of an incoming job using only the resource-utilization patterns observed prior to the jobs arrival. Our load indices combine information from the key resources of contention: CPU, disk, network, and memory.

Adaptive load-balancing strategies for distributed systems

Describes SMALL, a system for learning load-balancing strategies in distributed computer systems. The load balancing problem is an ill-posed optimization problem because its objective function is ill-defined. Realistic state-space representations of this problem do not satisfy the Markov property. Experimentally feasible learning environments for load balancing exhibit delayed, evaluative feedback. Such aspects complicate the learning of strategies for load balancing. SMALL uses comparator neural networks for learning to compare objective-function values of states resulting from a set of alternative moves. The problem of learning from delayed evaluative feedback, also called the credit-assignment problem of reinforcement learning, is solved only for Markovian problems. The paper presents a novel credit-assignment procedure suitable for load balancing and other non-Markovian learning tasks.<<ETX>>

Physical-level synthetic workload generation for load-balancing experiments

Synthetic workload generation uses artificial programs to mimic the resource-utilization patterns of real workloads. It is important for systematic evaluation of dynamic load-balancing strategies because load-balancing experiments require the measurement of task-completion times under realistic and reproducible workloads. The authors describe a generator that permits accurate replay of measured system-wide loads, thus providing an ideal setting for conducting load-balancing experiments. The generator is implemented inside the operating-system kernel and, therefore, has complete control over the local resources. It controls the useage levels of four key resources: CPU, memory, disk, and network. In order to reproduce accurately the behavior of the process population generating the measured load, the generator gives up a fraction of its resources in response to the arrival of new jobs, and reclaims these resources when the jobs terminate. The authors results show near-perfect reproduction of background load even in the presence of interfering foreground load.<<ETX>>

Load Balancing: An Automated Learning Approach

This book presents a system that learns new load indices and tunes the parameters of given migration policies. The key component is a dynamic workload generator that allows off-line measurement of task-completion times under a wide variety of precisely controlled loading conditions. The workload data collected are used for training comparator neural networks, a novel architecture for learning to compare functions of time series and for generating a load index to be used by the load balancing strategy. Finally, the load-index traces generated by the comparator networks are used in a population-based learning system for tuning the parameters of a given load-balancing policies. Together, the system constitutes an automated strategy-learning system for performance-driven improvement of existing load-balancing software.

Population-based learning of load balancing policies for a distributed computer system

Effective load-balancing policies use dynamic resource information to schedule tasks in a distributed computer system. In this paper, we present a novel method for automatically learning such policies. At each site in our system, we use a comparator neural network to predict the relative speedup of an incoming task using only the resource-utilization patterns obtainedprior to the tasks arrival. Outputs of these comparator networks are broadcast periodically over the distributed system, and the resource schedulers at each site use these values to determine the best site for executing an incoming task. The delays incurred in propagating workload information and tasks from one site to another, as well as the dynamic and unpredictable nature of workloads in multiprogrammed multiprocessors, may cause the workload pattern at the time of execution to differ from patterns prevailing at the times of load-index computation and decision making. Our loadbalancing policy accommodates this uncertainty by using certain tunable parameters. We present a population-based machine-learning algorithm that adjusts these parameters in order to achieve high average speedups with respect to local execution. Our learning algorithm overcomes the lack of taskspecific information by learning to compare the relative speedups of different sites with respect to the same site, rather than attempting to predict absolute speedups. We present experimental results based on the evecution of benchmark programs on a network of Sun workstations connected by a local area network. Our results show that our load-balancing policy, when combined with the comparator neural network for workload characterizatioa is effective in exploiting idle resources in a distributed computer system.

Automated learning of load-balancing strategies in multiprogrammed distributed systems

Abstract Dynamic load-balancing strategies for distributed systems seek to improve average completion time of independent tasks by migrating each incoming task to the site where it is expected to finish the fastest: usually the site having the smallest load index. SMALL is an offline learning system for developing configuration-specific load-balancing strategies; it learns new load indices as well as tunes the parameters of given migration policies. Using a dynamic workload generator, a number of typical systemwide load patterns are first recorded; the completion times of several benchmark jobs are then measured at each site, under each of the recorded load patterns. These measurements are used to train comparator neural networks simultaneously, one per site. The comparators collectively model a set of perfect load indices in that they seek to rank, at arrival time, the possible destinations for an incoming task by their (not yet known) respective completion times. The numerous parameters of the decentral...

Constructing representations using inverted spaces

This paper presents a novel technique for the construction and evaluation of new features for a two-class discrimination learning task. Our method employs statistically sound heuristics for guiding the knowledge-based construction of feature combinations. An attempt is made to integrate the strategies for construction of new features of description with the strategies for selecting a sufficient, representative training set.

STRATEGY LEARNING: A SURVEY OF PROBLEMS, METHODS, AND ARCHITECTURES

Problem solvers employ strategies in searching for solutions to given problem instances. Strategies have traditionally been designed by experts using prior knowledge and refined manually using trial and error. Recent attempts to automate these processes have produced strategy-learning systems. This paper shows how various issues in strategy learning are affected by the nature of performance tasks, problem solvers, and learning environments. Surveyed learning systems are grouped by the commonality of their approaches into four general architectures.