Queue-Mem: Energy-Efficient Hardware Storage for Advanced Network Function Acceleration
summary
Queue-Mem stores packet payloads in the switch shared packet buffer by pausing/resuming port queues with a batch-based queue control that maps five-tuple hashes to queue slices. Implemented on Tofino hardware and a single commodity server, it processes 1 Tb/s of 1.5 KB packets by handling ~100 Mpps of headers, keeps buffer use under 10 MB, latency under 30 µs, and TCP retransmits ≈1%. The paper reports energy improvements of 1.8×–10.9× over NIC/DRAM-based parking and shows robust operation across synthetic and real workloads.
AI-written summary · about
chapters (11)
- 0:00QMEM introduction and motivation
- 1:01Energy impact of terabit NFs
- 3:49Storing payloads in switch buffer
- 6:35Queue control strategies
- 7:58Pause/resume queue challenges
- 9:00Batch-based queue control in QMan
- 12:09QMEM implementation energy gap
- 12:30Evaluation with dynamic traffic
- 13:57Advanced NFs at terabit speeds
- 14:40Real-world TCP retransmissions
- 15:21Conclusion and future API
related talks
20mPostMan: Rapidly Mitigating Bursty Traffic by Offloading Packet Processing
16mAchieving Wire-Latency Storage Systems by Exploiting Hardware ACKs
22mProgrammable Calendar Queues for High-speed Packet Scheduling
32mSnabb Switch simple and fast packet networking
20mMemory Recycling for Network Interfaces
16mA High-Speed Stateful Packet Processing Approach for Tbps Programmable Switches
25mApproximating Fair Queueing on Reconfigurable Switches
13mClubHeap: A High-Speed and Scalable Priority Queue for Programmable Packet Scheduling