SIRD: A Sender-Informed, Receiver-Driven Datacenter Transport Protocol
summary
SIRD combines receiver-driven downlink scheduling with reactive control on shared links, implemented on Caladan and evaluated in a 144‑host simulator and a 100Gb testbed. Receivers start with 1.5×BDP default credit and allocate up to 1 BDP per sender while senders mark excess with a CSN bit so receivers run AI/MD and converge to about 0.35×BDP per congested sender. Results show matching Homa throughput, near‑optimal latency, and roughly an order of magnitude less queuing.
AI-written summary · about
chapters (14)
- 0:00SIRD and congestion control
- 1:18Congestion control importance
- 2:40Sender vs receiver approaches
- 4:27Limitations of current protocols
- 5:31SIRD design and credit allocation
- 6:58Incast handling with admission control
- 8:43Sender congestion and credit overcommit
- 10:48Reactive credit adjustment
- 12:12Core congestion with shared links
- 13:15Reducing message latency
- 14:23Evaluation setup and metrics
- 15:27SIRD vs sender/receiver baselines
- 15:49SIRD vs Homa
- 16:11Summary
related talks
14mUnderstanding RDMA Microarchitecture Resources for Performance Isolation
27msRDMA -- Efficient NIC-based Authentication and Encryption for Remote Direct Memory
23mMulti-Path Transport for RDMA in Datacenters
16mAn edge-queued datagram service for all datacenter traffic
15mReverie: Low Pass Filter-Based Switch Buffer Sharing for Datacenters with RDMA and TCP...
18mBURST: Seeking High-performance, Interoperability and Scalability in Soft-RDMA
16mRingLeader: Efficiently Offloading Intra-Server Orchestration to NICs
17mSRNIC: A Scalable Architecture for RDMA NICs