Predicting Resource Demand in Heterogeneous Active Networks

Recent research, such as the Active Virtual Network Management Prediction (AVNMP) system, aims to use simulation models running ahead of real time to predict resource demand among network nodes. If accurate, such predictions can be used to allocate network capacity and to estimate quality of service. Future deployment of active-network technology promises to complicate prediction algorithms because each “active” message can convey its own processing logic, which introduces variable demand for processor (CPU) cycles. This paper describes a means to augment AVNMP, which predicts message load among active-network nodes, with adaptive models that can predict the CPU time required for each “active” message at any activenetwork node. Typical CPU models cannot adapt to heterogeneity among nodes. This paper shows improvement in AVNMP performance when adaptive CPU models replace more traditional non-adaptive CPU models. Incorporating adaptive CPU models can enable AVNMP to predict active-network resource usage farther into the future, and lowers prediction overhead. INTRODUCTION Growing availability of processing power and bandwidth in communication networks encourages innovative approaches to network management. One specific innovative idea envisions injecting simulation models into network nodes, and then running those models in parallel with the operational network, but ahead in time, in order to predict traffic and resource use. If the models predict accurately, then network management systems can better allocate capacity in anticipation of varying demands and network operators can better estimate the quality of service (QoS) that customers can expect. This paper describes one approach, the Active Virtual Network Management Prediction (AVNMP) system [1], which aims to predict network traffic. AVNMP uses active-network technology [2] to inject simulation models into network nodes, and to run those models concurrently with corresponding applications. AVNMP then compares estimated performance against measured performance, and maintains predictions from the simulation within specified error bounds, when compared against measurements from the application. AVNMP can estimate resource requirements for each node in a conventional communication network, such as the Internet. In conventional networks, accurate estimates for message quantity and size directly imply nodal resource requirements for bandwidth, memory, and processor (CPU) cycles. This holds because packets receive the same fundamental processing within each node. Unfortunately, the future deployment of active network technology promises to negate the simple, fixed relationship between packets and resource use. This will occur because each packet in an active network can carry code, or a reference to code, which must be loaded on demand and applied to the packet. This implies that in active networks the processing of individual packets can differ, demanding varying quantities of node resources, particularly CPU cycles. We set out to investigate how AVNMP might be used to predict resource consumption in an active network. This paper reports our initial findings. The paper is organized into seven sections. First, we provide a brief tutorial on active networks. Second, we describe how AVNMP uses active-network technology to predict traffic load in a conventional network. Third, we discuss how we augmented AVNMP to predict CPU usage in heterogeneous active networks. Here heterogeneity implies that the active network comprises a wide range of node types with various hardware capabilities and software configurations. This creates additional complexity because the demand for CPU cycles varies not only by packet type but also by node type. Fourth, we outline an experiment where we used AVNMP to predict resource consumption by an active audio application. Our results suggest that adaptive CPU models [3], which accommodate variations in node capabilities, can improve the accuracy of AVNMP predictions, and reduce prediction overhead. Fifth, we suggest some additional applications for AVNMP, and similar prediction systems. Sixth, we identify some future research suggested by our work. Finally, we present our conclusions from the current experiment. 0-7803-7227-1/01/$17.00 (c) 2001 IEEE