Static Inference of Transmission Data Sizes in Distributed Systems

Abstract

We present a static analysis to infer the amount of data that a distributed system may ransmit. The different locations of a distributed system communicate and coordinate their actions by posting tasks among them. A task is posted by building a message with the task name and the data on which such task has to be executed. When the task completes, the result can be retrieved by means of another message from which the result of the computation can be obtained. Thus, the transmission data size of a distributed system mainly depends on the amount of messages posted among the locations of the system, and the sizes of the data transferred in the messages. Our static analysis has two main parts: (1) we over-approximate the sizes of the data at the program points where tasks are spawned and where the results are received, and (2) we over-approximate the total number of messages. Knowledge of the transmission data sizes is essential, among other things, to predict the bandwidth required to achieve a certain response time, or conversely, to estimate the response time for a given bandwidth. A prototype implementation in the SACO system demonstrates the accuracy and feasibility of the proposed analysis.