Search

Real-time problem diagnosis in large distributed computer systems and networks is a challenging task that requires fast and accurate inferences from potentially huge data volumes. In this paper, we propose a cost-efficient, adaptive diagnostic technique called active probing. Probes are end-to-end test transactions that collect information about the performance of a distributed system. Active probing uses probabilistic reasoning techniques combined with information-theoretic approach, and allows a fast online inference about the current system state via active selection of only a small number of most-informative tests. We demonstrate empirically that the active probing scheme greatly reduces both the number of probes (from 60% to 75% in most of our real-life applications), and the time needed for localizing the problem when compared with nonadaptive (preplanned) probing schemes. We also provide some theoretical results on the complexity of probe selection, and the effect of 'noisy' probes on the accuracy of diagnosis. Finally, we discuss how to model the system's dynamics using dynamic Bayesian networks (DBNs), and an efficient approximate approach called sequential multifault; empirical results demonstrate clear advantage of such approaches over 'static' techniques that do not handle system's changes. Index Terms--Bayesian networks (BNs), computer networks, diagnosis, distributed systems, end-to-end transactions, information gain, probabilistic inference.

In this paper, we consider distributed systems that can be modeled as finite state machines with known behavior under fault-free conditions, and we study the detection of a general class of faults that manifest themselves as permanent changes in the next-state transition functionality of the system. This scenario could arise in a variety of situations encountered in communication networks, including faults occurred due to design or implementation errors during the execution of communication protocols. In our approach, fault diagnosis is performed by an external observer/diagnoser that functions as a finite state machine and which has access to the input sequence applied to the system but has only limited access to the system state or output. In particular, we assume that the observer/diagnoser is only able to obtain partial information regarding the state of the given system at intermittent time intervals that are determined by certain synchronizing conditions between the system and the observer/diagnoser. By adopting a probabilistic framework, we analyze ways to optimally choose these synchronizing conditions and develop adaptive strategies that achieve a low probability of aliasing, i.e., a low probability that the external observer/diagnoser incorrectly declares the system as fault-free. An application of these ideas in the context of protocol testing/classification is provided as an example. Index Terms--Communication protocols, discrete event system (DES), distributed systems, finite-state machines (FSMs), Markov models, probabilistic fault detection.

The model used for real-time network security and transmission capability analyses requires the use of a large area network and has to be brought up to 'real-time' conditions. To build a complete network model, supervisory control and data acquisition (SCADA) data have to be made available to many external bodies. In this paper, an Internet interface to SCADA systems is proposed for inter control center data exchange. A discrete event technique is utilized to simulate the interarea data exchange. The performance modeling of data exchange on the Internet is established, and a quantitative analysis of the data exchange delay is presented. Simulation and experimental results have indicated that the Internet provides a convenient way for intercontrol center data exchange and, by using state estimation (SE) accurate complete network models, can be obtained for practical security and available transfer capability analyses. Index Terms--Communication network simulation, complete network model, intercontrol center data exchange, security analysis.

The shared memory in a LAN-like environment (SMILE) project at Lehrstuhl fur Rechnertechnik und Rechnerorganisation, Technical University of Munich (LRR-TUM) investigates in high-performance cluster computing using system area networks. In the context of this project, a hardware monitor is being developed to observe the system area network (SAN) traffic. This hardware monitor is, therefore, capable of delivering detailed information about the run-time communication behavior of applications running on SMILE clusters. The central part of this monitor consists of a content-addressable counter array managing a small working set of the most recently referenced memory regions. Index Terms--Cluster, scalable coherent interface, monitoring.

This paper reports recent work on ethernet traffic generation and analysis. We use gigabit ethernet network interface cards (NICs) running customized embedded software and custom-built 32-port fast ethernet boards based on field programmable gate arrays (FPGAs) to study the behavior of large ethernet networks. The traffic generation software is able to accommodate many traffic distributions with the ultimate goal of generating traffic that resembles the data collection system of the ATLAS experiment at CERN, Geneva, Switzerland. Each packet is time stamped with a global clock value and, therefore, we are able to compute an accurate measure of the network latency. Various other information collected from the boards is displayed in real time on a graphical interface. This work provides the tools to study a test bed representing a fraction of the 1600 ATLAS detector readout buffers and 600 Level 2 trigger central processing units (CPUs) using a combination of the fast ethernet boards and the gigabit ethernet NICs. Index Terms--Measurement, modeling, networks, switches.

A sampling method proposed by Fishman to compute 2-terminal and global reliability of a network is considered.

The two common techniques for testing the terminal-pair reliability within a computer network are PB and CB. However, a new technique has been developed based on cut-based algorithms. A comparison of the new technique and PB and CB was conducted on the basis of the number of subproblems solved and computation time. Results showed that the cut-based technique is superior to the two path-based techniques in terms of subproblem solving efficiency and computational time.

We will present results from the first year of operation of the STAR data-acquisition (DAQ) system using a Myrinet Network. STAR is one of four experiments to have been commissioned at the Relativistic Heavy Ion Collider (RHIC) at Brookhaven National Laboratory during 1999 and 2000. The DAQ system is fully integrated with a Level 3 Trigger. The combined system currently consists of 33 Myrinet Nodes, which run in a mixed environment of MVME processors running VxWorks, DEC Alpha workstations running Linux, and SUN Solaris machines. The network will eventually contain up to 150 nodes for the expected final size of the L3 processor farm. Myrinet is a switched high.speed low-latency network produced by Myricom and available for PCI and PMC on a wide variety of platforms. The STAR DAQ system uses the Myrinet network for messaging, L3 processing, and event building. After the events are built, they are sent via gigabit ethernet to the RHIC computing facility and stored to tape using HPSS. The combined DAQ/L3 system processes 160-MB events at 100 Hz, compresses each event to 20 MB, and performs tracking on the events to implement a physics-based filter to reduce the data storage rate to 20 MB/s. Index Terms--Computer network performance, computer networks, nuclear physics, SCADA systems.

The ATLAS second-level trigger will use a multilayered local-area network to transfer 5-Gbyte/s detector data from ~1500 buffers to a few hundred processors. A model of the network has been constructed to evaluate its performance. A key component of the network model is a model of an individual switch, reproducing the behavior measured in real devices. A small number of measurable parameters are used to model a variety of commercial Ethernet switches. Using parameters measured on real devices, the impact on the overall network performance is modeled. In the Atlas context, both 100-Mbit and gigabit Ethernet links are required. A system is described that is capable of characterizing the behavior of commercial switches with the required number of nodes under traffic conditions resembling those to be encountered in the Atlas experiment. Fast Ethernet traffic is provided by a high-density custom-built tester based on field-programmable gate arrays, programmed in Handel-C and VHDL, while the gigabit Ethernet traffic is generated using Alteon NICs with custom firmware. The system is currently being deployed with 32 100-Mbit ports and 16 gigabit ports, and will be expanded to ~256 nodes of 100 Mbit and ~50 GBE nodes, Index Terms-Field-programmable gate arrays (FPGAs), modeling, networks, network testing, physics, triggering.