Presentation Schedule for Analysis and Prediction...

The goal of WatchTower is to provide a maximally unobtrusive performance monitor for the Windows 2000 operating system. Towards this end, the utility has been built to run as a service that begins automatically upon system start up and uses a practically undetectable amount of system resources. Being a service, it does not present the user with distracting windows and it does not need to be manually started, stopped or configured once installed. As WatchTower runs it creates uniquely named log files on an hourly basis in the well-known comma separated format (CSV). These files contain measurements of counters (a performance data item related to a specific area of system functionality, such as a processor's busy time, memory usage, or the number of bytes received over a network connection), which can be later analyzed. In this paper, we describe the current system architecture of WatchTower and show why it is a useful addition to the list of existing performance monitoring tools for the Windows 2000 operating system.

Bulk Synchronous Programming Model is known for designing efficient and portable data parallel algorithms. The efficiency of a bulk synchronous program depends on a number of parameters. Bulk Synchronous programs can be implemented using various techniques. This paper will deal with bulk synchronous programs implemented using Message Passing Interface (MPI). Some of the very common bulk synchronous programs have been analyzed (to track their behavior). An effort has also been made to predict the exact behavior of the system. In doing this, various parameters that affect the behavior of bulk synchronous programs have been identified. These parameters, when carefully calibrated, can be used for improving the performance and throughput of bulk syncronous programs.

About a thousand odd parameters (for example, Memory\Available Bytes, Disk\Transfers/sec, Processor\Privileged_Time, etc.) are available on a Windows NT/2000 Workstation at a frequency of 1Hz via Perfmon. One day worth of all counters logs can be about 2GB. Such a wealth of information, but unmanageably large to store, maintain or analyze.

The behavior of these parameters tells us about the state of the workstation, and its dynamics. Previous work on estimating host loads makes a priori assumptions about what parameters are representative of the state of a workstation. Here I dismiss that assumption, and look at the statistical properties of the data to discover them.

I will present linear and nonlinear methods of dimensionality reduction, maybe talk a little about clustering, and how these applied to Windows logs; and results - the important dimensions, and their dynamics.

As wireless mobile devices becoming increasingly popular, it becomes evident that understanding the nature of such wireless systems is paramount to improving their performance. It is well known that the error rate for wireless communication is higher than its wireline counterpart. Most wireless research has focused on characterizing this loss behavior. This paper describes an effort to analyze a set of trace data collected from WaveLan II Access Points at Carnegie Mellon University. I look at four key dimensions : quantity and error rates for incoming and outgoing data over time. The analysis indicated there was strong autocorrelation both in the quantity of outgoing data and the incoming error rate. Surprisingly, no cross correlation was found between any pair of discrete dimensions. Finally, estimates of the Hurst Parameter show that the data sets appear to be long-range dependent.

In many network and computer performance measurement systems an elaborate amount of data is collected in order to provide a historic picture of the dynamics of a network and computer system on that network over a particular time period. In order to ease the storage constraints of these measurement systems, a wavelet based compression scheme is devised. A compression methodology will allow storage facilities to provide more measurement history to an application that can then query information over weeks instead of days. This form of compression, while lossy, should exhibit the dynamics of the original stream and a confidence interval should be provided so that the application understands the accuracy of the given query. In this talk, a weeklong measurement stream of host load traces are analyzed in order to prove the concept of compression using the wavelet approach.