[论文解读] Causality, Knowledge and Coordination in Distributed Systems

Ido Ben-Zvi|arXiv (Cornell University)|Dec 19, 2011

Distributed systems and fault tolerance参考文献 45被引用 7

一句话总结

本文通過形式化框架研究了分布式系統中的因果關係、知識與協調機制，該框架對嵌套知識與共同知識、事件排序以及無知知識進行建模。論文提出了一種廣義的事件排序機制，並證明實現共同知識可實現可靠的協調。關鍵結果顯示，無知知識可系統性地建模與利用，進而用於同步系統中的容錯協議。

ABSTRACT

Effecting coordination across remote sites in a distributed system is an essential part of distributed computing, and also an inherent challenge. In 1978, an analysis of communication in asynchronous systems was suggested by Leslie Lamport. Lamport's analysis determines a notion of temporal precedence, a sort of weak notion of time, which is otherwise missing in asynchronous systems. This notion has been extensively utilized in various applications. Yet the analysis is limited to systems that are asynchronous. In this thesis we go beyond by investigating causality in synchronous systems. In such systems, the boundaries of causal influence are not charted out exclusively by message passing. Here time itself, passing at a uniform (or almost uniform) rate for all processes, is also a medium by which causal influence may fan out. This thesis studies, and characterizes, the combinations of time and message passing that govern causal influence in synchronous systems. It turns out that knowledge based analysis [FHMV] provides a well tailored formal framework within which causal notions can be studied. As we show, the formal notion of knowledge is highly appropriate for characterizing causal influence in terms of information flow, broadening the analysis of Chandy and Misra in [ChM]. We define several generic classes of coordination problems that pose various temporal ordering requirements on the participating processes. These coordination problems provide natural generalizations of real life requirements. We then analyze the causal conditions that underlie suitable solutions to these problems. The analysis is conducted in two stages: first, the temporal ordering requirements are reduced to epistemic conditions. Then, these epistemic conditions are characterized in terms of the causal communication patterns that are necessary and sufficient to bring them about.

研究动机与目标

• 形式化知識（特別是嵌套知識與共同知識）在分布式程序之間實現可靠協調中的作用。
• 解決在部分資訊與非同步通訊環境下，實現一致全局狀態的挑戰。
• 將無知知識建模為檢測不一致與提升系統韌性的機制。
• 將事件排序廣義化，超越因果關係，以支援同步環境中的基於知識的協調。
• 為設計透過共享知識與時間排序實現協調的協議，提供邏輯基礎。

提出的方法

• 使用形式化的知識邏輯框架，對分布式代理之間的知識與共同知識進行建模。
• 基於因果依賴與知識傳播，提出廣義的事件排序機制。
• 應用「無知知識」概念，以檢測不一致並防止錯誤的協調決策。
• 採用分層方法，從個體知識逐步構建嵌套知識結構，直至達成共同知識。
• 利用同步系統假設，確保知識傳播在有界時間內收斂。
• 結合時序邏輯與知識邏輯，模擬知識如何透過訊息交換與系統事件演變。

实验结果

研究问题

• RQ1 如何在分布式系統中系統性地構建與維持嵌套知識？
• RQ2 代理達成共同知識所需滿足的條件為何？其如何促成協調？
• RQ3 如何利用無知知識檢測並防止分布式協議中的錯誤協調？
• RQ4 因果排序與分布式系統中的知識邏輯之間的關係為何？
• RQ5 廣義事件排序如何支援同步環境中基於知識的協調？

主要发现

• 在同步系統中，透過迭代式知識傳播可達成共同知識，進而實現可靠協調。
• 無知知識被形式化建模，並證明其對檢測不一致與防止錯誤決策至關重要。
• 嵌套知識結構可逐步建立，每一層次代表代理間更深入的理解。
• 廣義事件排序透過納入知識約束，擴展了 Lamport 的因果排序，提升了協調的可靠性。
• 該框架證明，共同知識對於解決共識與快照協議等協調問題而言，既是必要也是充分條件。
• 本文確立，基於知識的協調優於純粹基於因果關係的方法，尤其在存在部分資訊與通訊延遲的系統中。