Explore a conference talk on KRCORE, an innovative RDMA library designed for elastic computing with a microsecond-scale control plane. Learn about the challenges of creating RDMA connections and discover how KRCORE addresses these issues through virtualization of pre-initialized kernel-space connections and retrofitting of advanced RDMA dynamic connected transport. Delve into the key ideas behind KRCORE's ability to establish full-fledged RDMA connections within 10μs, significantly faster than traditional methods. Examine the implementation details, including connection pooling, QP sharing, and the MetaServer design. Gain insights into KRCORE's performance evaluations, its impact on disaggregated key-value stores, and its potential applications in serverless computing. Understand how this technology can dramatically reduce worker bootstrap time and data transfer latency in elastic computing environments.
KRCORE - A Microsecond-scale RDMA Control Plane for Elastic Computing
Overview
Syllabus
Intro
Remote Direct Memory Access (RDMA)
Problem: creating RDMA connections is slow
No problem for traditional applications
Impact the performance of elastic applications
Basic idea: connection pooling & reusing
User-space QPs cannot be shared
Solution #1. share QPs in a kernel-space QP poc
Massive QPs cached in the pool
Solution #2. Retrofit DCT as the shared connecti
Our design: MetaServer
Correct QP multiplexing
We add additional checks to prevent QP corrupt
Put it all together: KRCore
KRCore implementation
Evaluations setup
Control plane performance of KRCore
Data plane performance
Accelerating disaggregated RDMA-enabled KVS
Summary and discussion
Conclusion of KRCore
Taught by
USENIX
