Characterizing Group Communication Middleware for Real-time

This paper presents our current work in characterizing the behavior of a real-time dependable distributed system, which must exhibit predictable behavior under load and in the presence of partial failures. We focus on measuring the end-to-end properties of

Characterizing Group Communication Middleware for Real-time Distributed SystemsL. M. Feeney, P. Bernadat, F. Travostino The Open Group Research Institute 11 Cambridge Center, Cambridge, MA 02142 flmfeeney, bernadat, travosg@http://doc.guandang.netThis paper presents our current work in characterizing the behavior of a real-time dependable distributed system, which must exhibit predictable behavior under load and in the presence of partial failures. We focus on measuring the end-to-end properties of the middleware implementing the real-time process group service, specifically its membership and message latency. The paper also describes the tools and techniques we have developed, along with some of the practical issues that arise in instrumenting a real-time distributed system.

Abstract

A major focus of research at The Open Group Research Institute is the development of real-time dependable distributed systems. One of our goals is to provide our academic and industrial partners with con gurable, reusable frameworks for research and further development. Target applications include factory automation, sensor monitoring and combat systems, which require predictable behavior, even in emergency conditions. Under this program, we have implemented middleware for a real-time process group service (GIPC). In contrast to group communication services which emphasize throughput (e.g., ISIS 2]), GIPC is distinguished by its focus on real-time issues, con gurability, and predictable behavior. This paper describes our ongoing work characterizing GIPC behavior under load and in the presence of partial failures. Ensuring real-time guarantees at the application-level requires appropriate real-time behavior at all levels in the system. Moreover, because of our focus on con gurability, we expose many tunable parameters within the system. Experience with how these parameters interact and how to select appropriate values for di erent application requirements is essential if we expect others to adopt our work.This work was supported in part by the Defense Advanced Research Projects Agency (DARPA) and the Rome Laboratory of the Air Force Materiel Command (AFMC).

1 Introduction

Measuring a live real-time system presents issues that do not arise when working with a simulation. The instrumentation must be minimally intrusive, and must not be adversely a ected by the load or failures whose e ects we are trying to measure. Specialized hardware, such as a synchronized clock, dedicated CPU or bus snooper, can be used to address some of these issues. However, this becomes costly for more than a few nodes and runs counter to our preference for general-purpose, commodity hardware. Instead, our approach uses the real-time properties of the system to isolate the instrumentation. Following a brief description of the system, we specify the high-level properties to be characterized and the variables on which they depend. We describe the tools and instrumentation we have developed and repor

t on some preliminary measurements. Early measurements have con rmed some expected properties of the system and indicated some anomalies, emphasizing the importance of this sort of characterization. We note that our tests are often speci cally constructed to measure one aspect of the system and do not generalize well. We speculate that a characterization framework is important for making these sorts of measurements. While our modular design and emphasis on con gurability allow us to support many group communication paradigms and qualities of service, our current focus is on the powerful semantics of FIFO atomic broadcast and the virtual synchrony model 2]. GIPC provides FIFO atomic broadcast with safe delivery to the failure domain of the atomic broadcast engine of each group member. It also guarantees that group members order all messages and membership views with a global sequence number. By implementing these semantics in a real-time environment, we have created a very strong base on which to build real-time dependable applications. The GIPC middleware is policy-neutral; the application is expected to participate in fault management and quality of service negotiation via the GIPC API.

2 Component Technologies

This paper presents our current work in characterizing the behavior of a real-time dependable distributed system, which must exhibit predictable behavior under load and in the presence of partial failures. We focus on measuring the end-to-end properties of

Some of GIPC's real-time properties, i.e. those based on reservation and prioritization, derive from its implementation within the CORDS (Communication Objects for Real-time Dependable Systems) 9] framework. CORDS is an object-oriented communication framework based on the University of Arizona x-kernel 7]. CORDS provides a framework for composing a graph (stack) of micro-protocol objects. Messages are shepherded through the protocol graph via object invocation. CORDS' most signi cant extension to the x-kernel framework is the path abstraction 10], which uni es resource (bu er and thread) reservation and allocation along each communication channel through the protocol graph. This mechanism provides for controlled interference among CORDS tra c. For high priority trafc, such as GIPC liveness messages, paths can be used ensure that reserved bu ers are available to receive the incoming packets and that high priority threads are waiting to shepherd them though the protocol graph. The CORDS framework is implemented in the MK7 operating system, a portable real-time microkernelbased system derived from Mach 3.1 Key features include kernel preemption and xed-priority …… 此处隐藏：12807字，全部文档内容请下载后查看。喜欢就下载吧 ……