The Berkeley/Stanford Recovery-Oriented Computing (ROC) Project Winter 2003 Retreat

• Slides

  • ROC @ Berkeley Status - Dave Patterson

  • ROC @ Stanford Status - Armando Fox

  • Automatic Failure Path Inference for J2EE - George Candea

  • Benchmarking and Debugging Very Large Scale Network Systems - David Oppenheimer

  • Building Undo for Operators: An Undoable E-mail Store - Aaron Brown

  • Crash-Only Software - George Candea

  • Dynamic Path-Based Macro Analysis - Mike Chen

  • Failure Analysis of Internet Systems - Archana Ganapathi

  • Free the Replicas! - Andy Huang

  • Metrics and Techniques for Evaluating the Performability of Internet Services - Pete Broadwell

  • Middle-tier State Store - Ben Ling

  • Status and Future Directions of Pinpoint - Emre Kiciman

• Posters

  • First Steps Towards ROC for Storage Systems - Atsushi Ishikawa, Ryusuke Ito, Hitachi Ltd.
    Recent storage systems require very high level skills to manage, due to their complex functionality and redundancies. This tendency is especially remarkable as regards performance management and maintenance. Because of the chronic shortage and high employment cost of skilled staff, the result is an increase in total cost of ownership (TCO).  To address the causes of this problem, we focus on the following three points. First, we categorize problems in performance management, maintenance handling and security of current DAS (Direct Attached Storage), SAN(Storage Area Network) and near-future storage systems. Next, we suggest various ROC solutions for these problems, which are also effective for TCO reduction. Finally, we suggest a basic approach to performance management and maintenance handling, which we expect to remain the biggest problem for real network-based storage environments in the near future.
  • Benchmarking and Debugging Large-Scale Decentralized Systems - David Oppenheimer
    Decentralized systems require new benchmarks and new benchmarking techniques. We describe a general methodology for benchmarking the performability of one class of decentralized system: peer-to-peer applications built on top of distributed hash tables (DHTs). Furthermore, we argue that benchmarks for decentralized systems must be designed and implemented with similar concern for scalability and robustness as the systems they are designed to benchmark, and that this requires decentralized load generation, fault injection, and metric collection. These criteria lead us to propose a benchmark design that uses a separate DHT to publish the faultload description and to store collected metrics, and uses a DHT-based relational query engine to analyze benchmark results. Finally, we describe how the fault injection and monitoring mechanisms required to run such benchmarks are reusable for online robustness testing, problem detection, and system-level debugging.

  • Decoupling Storage: "Free the Replicas!" - Andy Huang, Armando Fox
    The goal of Decoupling Storage (DeStor) is to design an application-level persistent storage system for Internet services. This poster describes: 1) the coupling that exists in many replication schemes (e.g., read-one-write-all), 2) why coupling is a key hinderance to building systems that exhibit good performance degradation and keep data available throughout recovery, and 3) a quorum-based system that uses dynamic load-balancing information to replicate data in a decoupled manner.

Last Updated: 02/12/2004 09:21