Building Scalable Data Centers: BGP is the Better IGP

 

Building Scalable Data Centers: BGP is the Better IGP

Petr Lapukhov, Microsoft

Design choices, such as protocols & topology, for large scale (100k+ services) Web data centers have significant impact on scalability and operations of underlying networks. Such networks often host Hadoop-like compute clusters, which calls for topology providing large amount of bisection bandwidth. Consequently, to satisfy the requirement for intense east-west traffic, operators commonly choose some of Clos/Fat-tree topology variants. Another prominent property of such environments is reliance on fixed form-factor, cheap commodity 1G/10G Ethernet switches to reduce capital expenditures. Both of these properties combined result in a network topology with very high link density and large number of switches – on the order of 10’s of thousands and thousands of elements respectively. Applying traditional data-center designs with large L2 domains, even leveraging features such as vPC/mLAG is not possible due to scalability issues. This situation calls for different design options

Building Scalable Data Centers: BGP is the Better IGP (pdf)