クラスタシステムに対する多重障害からの回復技術: 歪チェックポイントモデルの厳密解法

摘要

This paper discusses the stochastic model for measuring software service availability; this is one of the customer-oriented attribute and denned as the attribute that the software system can successfully satisfy the end users' requests. From the viewpoint of a user, occurrence of a system failure is recognized when either one of the following two events arises: a software failure occurs when the user is using and operating the system, or a usage request occurs when the system is down. We propose and derive three kinds of novel service-oriented software availability assessment measure. The time-dependent behaviors of the system alternating between up and down state and the user's request are described by a Markov process. Then we incorporate the operation-oriented restoration scenario into the model, i.e., we consider the following two types of restoration: one is the restoration with debugging and the other is without debugging. Furthermore, the dynamic software reliability growth process, the upward tendency in difficulty of debugging, and the imperfect debugging environment are also included in software service availability modeling. Finally, we present several numerical examples of these measures for software service availability analysis.%計算機クラスタには複数ノードが存在するため，各単独ノードに対する障害だけでなく複数ノードに対する多重障害を考慮しなければならない・本稿ではNakamura．（2004）によって提案された歪チェックポイント（Skewed Checkpoint）モデルの再定式化を行い，重み付き期待オーバーへ、？ドを最小にする最適チェックポイント間隔を導出するための厳密解法を提案する．全ての多重障害の発生可能性を視覚的に列挙し，クラスタシステムの取り得る確率的事象を記述するためのグラフモデルを新たに提案し，重み付き期待オーバーヘッドの正確な表現を得るための形式的方法論について述べる．最終的に数値例を通じて，最適な盃チェックポイント間隔のモデルパラメータに関する感度分析を行い，クラスタシステムに対する環境分散技術の効果について調べる．