SPDK Based User Space NVMe Over TCP Transport Solution
summary
SPDK implements both host and target NVMe-over-TCP transports, interoperates with the Linux kernel, and shows concrete benchmarks: 4K random 70/30 R/W workloads with up to ~16 initiator connections, delivering roughly 2–2.5× better throughput for the same CPU and about 20–30% lower latency versus kernel targets. Implementation uses a polled user-space threading model and a socket wrapper; future work includes user-space TCP (VPP+DPDK) integration and ADQ on Intel E810 NICs to further cut latency and offload CPU.
AI-written summary · about
chapters (12)
- 0:00Agenda
- 0:50Why SPDK
- 11:57SDK
- 20:08Efficiency Stack
- 20:51Read Command
- 22:50Performance
- 25:09Latency
- 26:41Latency Comparison
- 28:43Ongoing Process
- 31:16Interoperability
- 32:03Conclusion
- 34:06Questions
related talks
39mSecured I/O in the Cloud with NVMe-Over-TCP
20mBuilding high efficient storage infrastructure for secure container on top of SPDK
22mBoosting the Network Performance of Confidential VM Using Userspace Stack
30mAccelerate TLS with DPDK
36mCryptodev and Security Library Updates in DPDK
21mHigh-Performance Cloud Native Traffic Authentication Solutions
54mThe Linux Network Stack
17mHarmonic: Hardware-assisted RDMA Performance Isolation for Public Clouds