Overview
Explore a conference talk on Tiara, a scalable and efficient hardware acceleration architecture for stateful Layer-4 load balancing. Learn about the challenges of distributing terabits per second of Internet traffic at datacenter boundaries and how Tiara addresses these issues. Discover the innovative three-tier architecture that utilizes heterogeneous hardware, including programmable switches and FPGAs for the fast path and x86 servers for the slow path. Understand how Tiara divides the load balancing fast path into memory-intensive and throughput-intensive tasks, mapping them to the most suitable hardware. Gain insights into the implementation of a fully functional Tiara prototype and its impressive performance metrics, including 1.6 Tbps throughput, 80M concurrent flows, and 1.8M new connections per second. Delve into the system goals, architecture details, traffic handling, optimizations, and comparative advantages over existing approaches in this comprehensive presentation from NSDI '22.
Syllabus
Intro
L4 LB at datacenter boundary
Being stateful
Stateful L4 LB requirements
Existing solution: software-based LB
Existing solution: switch-based LB
Strawman solution: switch-server LB
Traffic at datacenter boundary
System goals
Tiara idea
Tiara three-tier architecture
Tiara architecture in details
Inbound traffic: the first packet
Inbound traffic: the rest packets
Outbound traffic
Optimizations
Prototype implementation T-switch: Barefoot Tofino switch
System performance
Latency-bounded throughput
Tiara vs. existing approaches
Conclusion
Taught by
USENIX