Enabling Programmable Infrastructure for Multi-Tenant Data Centers

Abstract

Today’s data centers are large, shared infrastructures hosting hundreds of thousands of tenants over a vast network of servers. Operating these data centers frequently entails incorporating new infrastructure services that require customizing the behavior of its switches and exploiting unique characteristics of data centers to scale.

Emerging programmable switch ASICs allow network operators to customize the behavior of physical switches. Yet, virtual switches running on servers in multi-tenant data centers are still fixed function and composed of large, complex software codebases. Modifying these switches requires both intimate knowledge of the switch codebase and extensive expertise in network protocol design, raising the bar for customizing these switches prohibitively high.

In this dissertation, we address these challenges by, first, presenting the design and implementation of PISCES: a programmable, protocol-independent software switch derived from Open vSwitch (OVS), a fixed-function hypervisor switch, whose behavior is customized using P4. PISCES is not tethered to specific protocols; this independence makes it easy to add new features. We also show how the compiler can analyze the high-level specification to optimize forwarding performance. Our evaluation shows that PISCES performs comparably to OVS and that P4 programs for PISCES are about 40 times shorter than equivalent changes to OVS source code.

Next, we demonstrate how such programmable switches help build scalable infrastructure services by exploiting the unique characteristics of data-center networks. We use these switches to address the multicast scalability problem in multi-tenant data centers and present the design of an infrastructure service, Elmo, that scales multicast by taking advantage of the data-center characteristics; specifically, the symmetric topology and short paths in a data center. In Elmo, a PISCES switch encodes multicast group information inside packets themselves, reducing the need to store the same information in hardware switches, which instead read the encoded information to route packets to recipients. In a three-tier data-center topology with 27,000 hosts, Elmo supports a million multicast groups using a 325-byte packet header, requiring as few as 1,100 multicast group-table entries on average in hardware switches, with a traffic overhead as low as 5% over ideal multicast.