This paper presents congestion control mechanisms for reliable and scalable message-oriented middleware following the publish/ subscribe communication model. We identify the key requirements of congestion control in this environment, how it differs from congestion control for the Internet, and propose a combination of two congestion control mechanisms, (1) driven by a publisher hosting broker (PDCC), (2) driven by a subscriber hosting broker (SDCC). SDCC decouples the notion of a receive window and a NACK window, and is used by subscriber hosting brokers in recovery mode. PDCC implements a scalable and low latency feedback loop between a publisher hosting broker and all subscriber hosting brokers, which is used to adjust the rate of publishing new messages, to allow brokers in recovery to eventually catch up, and other brokers to keep up. We present a detailed experimental evaluation of our implementation of these mechanisms in the Gryphon system by injecting network failures and link congestion.
展开▼
