Imad Antonios

Hierarchical performance modeling for distributed system architectures

Performance modeling and evaluation techniques are essential when designing and implementing distributed software systems. Constructing performance models for such systems can require significant effort. This paper presents Hierarchical Performance Modeling as a technique to model performance for different layers of abstraction. Once the system architecture and software functionality have been specified, this model supports performance model generation for the evaluation and analysis of computation delays of software processes, communication delays of distributed software architectures, and hardware platform alternatives. A simplified example is presented to illustrate the concepts of the Hierarchical Performance Model.

Performability Models for Multi-Server Systems with High-Variance Repair Durations

We consider cluster systems with multiple nodes where each server is prone to run tasks at a degraded level of service due to some software or hardware fault. The cluster serves tasks generated by remote clients, which are potentially queued at a dispatcher. We present an analytic queueing model of such systems, represented as an M/MMPP/1 queue, and derive and analyze exact numerical solutions for the mean and tail-probabilities of the queue-length distribution. The analysis shows that the distribution of the repair time is critical for these performability metrics. Additionally, in the case of high-variance repair times, the model reveals so-called blow-up points, at which the performance characteristics change dramatically. Since this blowup behavior is sensitive to a change in model parameters, it is critical for system designers to be aware of the conditions under which it occurs. Finally, we present simulation results that demonstrate the robustness of this qualitative blow-up behavior towards several model variations.

An analytic performance model of parallel systems that perform N tasks using P processors that can fail

We present a Markov model for analyzing the performance of parallel/distributed processors that execute a job consisting of N independent tasks in parallel using P processors. The model is a Markov chain with states representing service and failure rates with k (0<k/spl les/P) active processors. The task-times and processor failures are both exponentially distributed. We derive a number of expressions to determine the mean execution time, probability of success, work, and other measurable quantities, all conditioned on the job finishing successfully. A prototype, implemented using an extended version of ACMPI, is used for actual experiments that are based on simulated task-times and processor failures. We present our results comparing the analytic model with the prototype for a range of values of processor failure rates. We also discuss extensions of the model and issues related to communication costs, approximations and effect of task-time distributions.

On the relationship between packet size and router performance for heavy-tailed traffic

The problem of characterizing the relationship between packet size and network delay has received little attention in the field. Research in that area has been limited to either simulation studies or empirical observations that are detached from analytic traffic modeling. From a queuing viewpoint, it is simple to show that these three variables are inter-related, which necessitates a more careful study. We present a traffic model of a router fed by ON/OFF-type sources with heavy-tailed burst sizes. The traffic model considered is consistent with the evidence that Web traffic is heavy-tailed. The analysis cases that are considered establish a quantitative characterization of the complex relationship among packet payload and header sizes, traffic burstiness, and router queuing delay.

A Performance Model of Gossip-Based Update Propagation

We consider the problem of propagating an update to nodes in a distributed system using two gossiping protocols. The first is an idealized algorithm with static and dynamic knowledge of the system, and the second is a simple randomized algorithm. We construct a theoretical model that allows us to derive work and completion time statistics under varying transmission delay distributions. Numerical results are obtained for both exponential and nonexponential transmission times using linear-algebraic queueing theory techniques. Additionally, we present the results of simulation experiments showing that under node churn assumptions, the randomized algorithms performance is qualitatively different than in a fault-free system.

Understanding the relationship between network traffic correlation and queueing behavior: A review based on the N-Burst ON/OFF model

Abstract Understanding the impact of network traffic properties on performance behavior in bottleneck links or larger networks is of primary interest to traffic analysts and network designers. Among the contributing factors, variance and correlation properties have been thoroughly studied and a large set of individual results have been obtained. However, these individual contributing factors are not sufficient to predict performance behavior. In this paper we review a unifying and versatile class of ON/OFF models through which the relationship among these parameters can be characterized and their influence on network performance be understood. The analytic performance results from the model show that there is a radically different queueing behavior when the ON period duration follows truncated power-tail distributions (even if truncated), as opposed to model variants where these distribution types are used for the OFF periods. All these models create correlation functions that only decay slowly. This motivates the development of a simple data analysis scheme to distinguish performance relevant correlation. The scheme is described both for interarrival and count processes of traffic data and its effectiveness is shown using real data traces.

Asynchronous gossip-based data propagation protocol

Data dissemination protocols govern interaction and exchange of data among nodes in a distributed system. An understanding of data transfer protocols provides insight into efficient middleware management. Due to their simplicity, scalability and fault-tolerance, gossip-based protocols are researched widely as an effective communication strategy. The Shuffle protocol presented in [1], is an example of a decentralized, gossip-based data transfer protocol used to spread information in a wireless network via probabilistic exchange of data. This paper presents, an asynchronous variant of the Shuffle protocol and a system model that captures variability in data transmission times. This transmission time variability is inherent in dynamic networks, where such algorithms are typically deployed. A simulation-based analysis of the protocols performance behavior is presented. Results show the effects of transmission variability, on data replication and its coverage. Also examined is the relationship between available storage and the performance of the protocol, expressed using measures such as propagation time and work.

A performance model and analysis of heterogeneous traffic with heavy tails

Several research efforts have recently attempted to characterize the effects of heavy-tailed traffic on router performance, while the effects of light-tailed traffic have for long been understood. Since general Web traffic originates from heterogeneous sources, the study of traffic mixing is of importance as it can reveal the degree to which heavy-tailed traffic can be handled by networking infrastructure before router delay becomes unacceptable. We present a model for heterogeneous traffic sources, where each is an ON/OFF process with an exponential OFF time and an arbitrary ON-time distribution. We consider the case of a 2-source process where one has exponential ON time and the other power tailed and present some analysis results showing the significance of traffic mixing on router performance.

A generalized analytic performance model of distributed systems that perform N tasks using p fault-p

A family of Markov models for analyzing the performance of parallel processors that execute a job consisting of N independent tasks using P fault-prone processors is presented in this paper. This study extends our previous study by allowing idle processors to fail, and also by developing performance models to analyze the case where one processor is fail-safe. The models are based on Markov Chains with states representing service, and failure rates with k (0 < k ? P) active processors. The task-times and processor failures are exponentially distributed. For each performance model we derive a number of expressions to determine the mean execution time, probability of success, standard deviation, work, and the average number of processor failures, all conditioned on the job finishing successfully. We present results of the models for a range of values of processor failure rates. In particular we analyze the results for N = P and N ? P.