PeRF: Preemption-enabled RDMA Framework
summary
PeRF implements RNIC preemption in software using managed QPs and zero-weight work requests to break large RDMA transfers into interleaved sub-requests and to pause/resume QPs via a preemption control QP. Prototype runs on a ConnectX‑6 100Gbps NIC with a 16‑core i7 and 32GB RAM, offloads control to a single dedicated core. Paper shows RDMA without isolation causes MF throughput to fall 3.8× and DF latency to rise 6.1×; PeRF matches hardware priority schemes for latency and outperforms token-based software (Justicia) on throughput and utilization.
AI-written summary · about
chapters (14)
- 0:00PeRF and RDMA intro
- 1:04Performance isolation issues
- 3:34Apps by message/QP levels
- 5:27RDMA hardware limits
- 6:30PeRF solution overview
- 7:13Preemption with managed QPs
- 9:45Message-level preemption
- 10:49QP-level pause/resume
- 12:15PeRF architecture
- 14:05PeRF evaluation setup
- 15:09Performance vs SP and ETSS
- 16:16Message isolation vs Justicia
- 18:04QP isolation in apps
- 19:32Conclusion
related talks
14mODRP: On-Demand Remote Paging with Programmable RDMA
28mPersistent Memory plus RDMA, new age remote device
12mPreview: Networking
14mUnderstanding RDMA Microarchitecture Resources for Performance Isolation
18mRDMA is Turing complete, we just did not know it yet!
15mFlor: An Open High Performance RDMA Framework Over Heterogeneous RNICs
17mWhite-Boxing RDMA with Packet-Granular Software Control
18mBURST: Seeking High-performance, Interoperability and Scalability in Soft-RDMA
