Backdraft: a Lossless Virtual Switch that Prevents the Slow Receiver Problem
Talkat 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22) · 2022-04-04..04-06·
summary
Lossless virtual switching implemented on BESS combines Dynamic Per-Flow Queuing to avoid head‑of‑line blocking, Doorbell queues to cut CPU overhead, and an overlay to prevent congestion spreading. Evaluation on a 100 Gbps cluster with DCTCP and Homa shows up to 20× lower 99th‑percentile tail latency versus lossy baselines. Key measured facts: memcached needs ~32 cores to reach line rate, nginx ~16 and iperf3 ~6; throughput swings of ~40 Gbps over ~100 μs; Homa with one core sees ≈10% drop rate.
AI-written summary · about
chapters (21)
- 0:00Intro
- 0:21Packet loss is a problem
- 0:31Packet loss can occur in the network
- 1:00Vagaries of CPU performance
- 1:46The slow receiver problem
- 2:11Lossless virtual switching is challenging
- 2:23Backdraft: A 10,000 Ft. View
- 2:49Insights of Backdraft
- 3:54Slow receivers manifest at us-scale
- 6:06Three components of Backdraft
- 6:32Key idea behind the per flow queuing
- 6:44Why do we lose packets?
- 7:50Design space of the Per Flow Queuing
- 8:28Dynamic Per Flow Queuing
- 9:55Key idea behind doorbell queues
- 10:07How does doorbell queues work?
- 11:51Backdraft overlay network
- 13:05Homa experiment setup
- 13:22Backdraft complements Homa - RPC Completion time
- 14:32Evaluating different components of
- 16:47Backdraft Takeaways Slow receivers are pervasive
related talks
23mBalancing on the Edge: Transport Affinity without Network State
32mLive Q+A 2
16mCredence: Augmenting Datacenter Switch Buffer Sharing with ML Predictions
20mHarmony: A Congestion-free Datacenter Architecture
18mRevisiting Congestion Control for Lossless Ethernet
15mReverie: Low Pass Filter-Based Switch Buffer Sharing for Datacenters with RDMA and TCP...
24mEnhancing Homa Linux for Efficient RPC Transportation
17mRe architecting Congestion Management in Lossless Ethernet