Patrick Lee and Hans-Arno Jacobsen.
MSRG, University of Toronto, July 2010.
Improvement in datacenter networks will enable architects to built larger distributed systems. However, standard distributed system architectures and their resulting communication patterns are incapable of supporting systems much larger than the current limits. This paper describes the design and evaluation of an overlay topology and messaging infrastructure that can be used for creating datacenter-scale distributed systems. The topology provides sufficient scalability to accommodate future datacenter growth. Theoretical analysis suggests its fault resiliency and network capability are unmatched by other common network topologies. The messaging infrastructure build on top of this topology supports incremental scaling, node failure recovery and flexible routing. Properties of the topology ensure that the overlay has constant bounded delay and bandwidth usage, which are useful for planning incremental deployment while guaranteeing performance. Finally, evaluation of our overlay prototype reveals usage trends that best matches the characteristics of our overlay
Readers who enjoyed the above work, may also like the following:
|