Brief announcement: sources of instability in data center multicast

Abstract

Data centers, and particularly the massive ones that support cloud computing, e-commerce, social networking and other large-scale functionality, necessarily replicate data. Our basic premise is that since updates to replicated data can be thought of as reliable multicasts, data center multicast is a potentially important technology. Nonetheless, a series of recent keynote speeches at major conferences makes it clear that data center multicast is a troubled area [4, 2]. One might expect such technologies to use IP multicast hardware, but in fact this is rare. Only TCP is really trusted today (because it backs down when loss occurs), and indeed, TCP is the overwhelming favorite among data center transport protocols [3]. Using TCP to get reliable multicast with high throughput produces an implicit TCP overlay tree.