Explore the groundbreaking discovery of Remote Direct Memory Access (RDMA) Turing completeness in this 18-minute conference talk from NSDI '22. Dive into RedN, a novel approach for implementing complex RDMA offloads without hardware modifications, using self-modifying RDMA chains. Learn how this technique lifts the existing RDMA verbs interface to a Turing complete set of programming abstractions, enabling offload of complex tasks like key lookups in distributed systems. Discover the potential performance improvements, including reduced latency for key-value get operations, enhanced performance isolation, and failure resiliency. Gain insights into the evaluation of RedN, its application in Memcached lookups, and the significant latency reductions achieved in both normal and contentious scenarios.
Overview
Syllabus
Intro
Benefits of RDMA networking
Massive growth in RDMA processing po
Existing designs for RDMA-based syste Commodity RNIC offloads
Alternative Design: Exploit RNIC Processing PL
Alternative Design: Exploit RNIC Processing Power
Conditional Branching - is it possible?
Branching with Self-Modifying Code
What about loops?
RedN Framework - Overview
Evaluation
Use case: Memcached Lookups
Results: Memcached get latency
Results: Memcached get contention
Conclusion
Taught by
USENIX
