HotDep ’06: Second Workshop on Hot Topics in System Dependability

S) Summarized by Geoffrey Lefebvre Making Exception Handling Work Bruno Cabral and Paulo Marques, University of Coimbra, Portugal Presented by Bruno Cabrel Exceptions are the standard mechanism for error handling in modern programming languages. Unfortunately, dealing with exceptions is a tedious process. Programmers often avoid the issue by writing empty handlers to save time. Programmers who take the effort to write proper exception-handling code see their productivity impaired. The authors argue that exception handling should not interfere with normal programming tasks but instead become a system issue. In this scenario, the run-time environment provides a set of generic exception handlers and deals with exceptions automatically. The programmer only has to wrap the code with try blocks at the appropriate locations. Because the system may have to try multiple handlers when an exception occurs, this approach requires that try blocks be resumable and appear atomic. Speculations: Providing Faulttolerance and Recoverability in Distributed Environments Cristian Tapus and Jason Hickey, California Institute of Technology Presented by Cristian Tapus Cristian began his talk by noting that distributed systems are ubiquitous. It is now mandatory that we build safe and reliable systems. Failures are frequent in highly parallel machines. It is critical that systems expected to run for a long time support fault tolerance. Unfortunately, traditional checkpoint mechanisms are application-specific, their implementation is expensive in terms of man-hours, and they are also error-prone. To address these issues, the authors present a new programming model based on speculative execution. The approach separates fault recovery code from computation. Fault recovery becomes transparent and automated and the design of distributed systems is simplified. Speculations are implemented as an extension to the Linux kernel. The implementation provides system calls to begin, abort, and commit speculative executions. Outbound messages sent while executing speculatively are marked accordingly. A process automatically switches to speculative execution when it receives a message marked as speculative. Discrete Control for Dependable IT