A software solution consists of multiple autonomous computations (i.e., execution threads) that execute concurrently (or apparently concurrently) over one or more locations to achieve a specific goal. Centralized solutions execute all computations on the same lo- cation while decentralized solutions disperse computations across different locations to increase scalability, enhance performance and reliability. Every location affects its executing computations both directly (e.g., the lack of a resource may prohibit a computation from progressing) and indirectly (e.g., an over- loaded location may slow down a computation). In a distributed environment, application developers have the luxury of executing each computation over its best-fitting location; the location (a) upon which the computation can achieve the best performance and (b) which guarantees the computation’s livelihood. Ideally, the decision to execute a computation over a location instead of another also load-balances the use of available resources such that it has the least impact over other computations (e.g., a computation should not execute over an already overloaded location further slowing down its computations). Application developers can only execute computations over their best-fitting location if their distributed programming language provides abstractions that allow them to control the locality of computations both before they are started and during their execution. In the rest of this document, section 2 briefly justifies why these two forms of locality control are required and section 3 outlines the issues that arise, and will be tackled in the talk to be held at CSAW 2014, by them.
展开▼
