Live in the Express Lane
summary
X-Lane isolates coordination traffic on a dedicated core and NIC queue, implemented as a kernel module and offloaded to a Metronome NFP4000 smartNIC, yielding single-digit microsecond latencies and nanosecond jitter. A heartbeat detector achieves detection under 10 μs, a Raft implementation runs at latencies under 20 μs, and replacing Redis replication produces 18x lower latency and 1.5x higher write throughput. DPDK and q-jump show double-digit millisecond outliers.
AI-written summary · about
chapters (10)
- 0:00Intro
- 0:23Crucial challenges of coordination tasks
- 3:03Features of X-Lane
- 3:50Separation between X-Lane & regular system
- 5:13Usual workflow for a packet
- 7:33X-Lane workflow for commodity hardware
- 9:29Latency and jitter for DPDK, QJump and X-Lane
- 10:34Tail latency over 21 days
- 11:25Latency and throughout results for Raft implementations
- 12:21Conclusion
related talks
16mRingLeader: Efficiently Offloading Intra-Server Orchestration to NICs
25mShenango: Achieving High CPU Efficiency for Latency-sensitive Datacenter Workloads
28mLow-Latency, Deterministic Networking with Standard Linux using XDP Sockets
28mRock around the clock (synchronization
18mKeepON: Supporting Deterministic Traffic on Standard NICs
19mAI/ML Networking Challenges: The Fast and the Finicky!
43mLessons Learnt from Reusing QDMA NIC to Base Band PMD
45mFaster or better designed? Choose any two!