Learnings from Deploying Network QoS Alignment to Application Priorities for Storage...
summary
Deploys an RPC-level mapping from application priorities to DSCP/WFQ classes (Aquidus) across storage traffic, performed via per-customer, per-region canaries sampling a fraction of traffic. Builds control vs. treated comparisons and telemetry for RPC latency, retries, timeouts and queue load. Finds per-RPC DSCP alignment fixes priority inversion, achieves 80% of request/response bytes aligned for four large Spanner clients, and shows downgrades can improve latency under overloaded queues.
AI-written summary · about
chapters (8)
- 0:00Network QoS for storage
- 1:42Network congestion effects
- 3:49RPC priorities and QoS costs
- 7:01QoS deployment challenges
- 10:12Spanner priority misalignment
- 11:59Production results after QoS
- 13:02Key learnings
- 15:09Conclusion and future
related talks
16mTowards Optimal Rack-scale μs-level CPU Scheduling through In-Network Workload...
17mBridging the Gap between QoE and QoS in Congestion Control: A Large-scale Mobile...
16mSöze: One Network Telemetry Is All You Need for Per-flow Weighted Bandwidth Allocation...
19mTackling DC Congestion and Bursts
25mRDMA @Scale
21mR2P2: Making RPCs first-class datacenter citizens
25mResQ: Enabling SLOs in Network Function Virtualization
19mNet-P4ct: Enhanced WAN Bandwidth Fair Sharing Using P4 Programmable Switches
