The combination of two inexorable trends---increasing parallelism and
increasing
big data analytics---means modern data centers urgently need efficient
high-capacity networks interconnecting thousands of servers. But we have
lacked
a fundamental understanding of what network topologies achieve high
throughput.
This talk shows how surprisingly, today's carefully-structured designs such as
fat-trees are far from optimal. Our Jellyfish architecture uses a completely
random network to achieve substantially higher throughput than deployed
designs, and
allows easy incremental expansion.  This leads to numerous important
theoretical
and practical questions in network design, such as optimality and the design
of
heterogeneous networks.