OpenSSL Crypto PMD - Analysis and Optimisations
summary
Analysis of the OpenSSL crypto PMD in DPDK shows the PMD spends <4% of time in crypto kernels versus a64 crypto at 82%, with atomic refcount ops consuming ~52% of runtime. Changes add per-queue cipher context clones (shipped in DPDK 24.07) to remove refcount thrash and an assembly interleaved AES-GCM port from a64 to OpenSSL (PR upstream), yielding double‑digit speedups on small buffers and ~5% on large single‑core runs, though OpenSSL still lags a64 below ~750‑byte buffers.
AI-written summary · about
chapters (6)
- 0:00Introduction and background
- 3:56OpenSSL PMD performance issues
- 9:12Optimizing cipher context handling
- 12:42Interleaved AES-GCM assembly
- 19:20Future optimizations and results
- 23:17Discussion and Q&A
related talks
36mCryptodev and Security Library Updates in DPDK
27mSpeeding Up Linux Disk Encryption
29mAre Our Systems Using up-to-Date Cryptography?
32mPQC in OpenSSL Library
37mOpenSSL Hardware Offload Enhancement
28mModernizing Kernel Cryptography: From Complex APIs To Streamlined Libraries
37mOpenSSL Hardware Offload Enhancement
47mLinux Cryptographic Acceleration on an i.MX6