mmb flexible high speed userspace middleboxes
play

mmb : Flexible High-Speed Userspace Middleboxes Korian Edeline , - PowerPoint PPT Presentation

mmb : Flexible High-Speed Userspace Middleboxes Korian Edeline , Justin Iurman, Cyril Soldani, Benoit Donnet Montefiore Institute, University of Lige Belgium A middleboxed Internet https://github.com/mami-project/roadshows 7/22/19 2 ANRW


  1. mmb : Flexible High-Speed Userspace Middleboxes Korian Edeline , Justin Iurman, Cyril Soldani, Benoit Donnet Montefiore Institute, University of Liège Belgium

  2. A middleboxed Internet https://github.com/mami-project/roadshows 7/22/19 2 ANRW 2019

  3. kernelspace vs userspace Kernel: Userspace: 3 ANRW 2019

  4. kernelspace vs userspace Kernel: ✗ T oo slow for high-speed forwarding ✗ Missing optimizations (batching, caching, etc) Userspace: 4 ANRW 2019

  5. kernelspace vs userspace Kernel: ✗ T oo slow for high-speed forwarding ✗ Missing optimizations (batching, caching, etc) Userspace: ✗ No direct access to NIC (context switch, sk_bufg) 5 ANRW 2019

  6. kernelspace vs userspace Kernel: ✗ T oo slow for high-speed forwarding ✗ Missing optimizations (batching, caching, etc) Userspace: ✗ No direct access to NIC (context switch, sk_bufg) ✔ DPDK (DMA, I/O batching) 6 ANRW 2019

  7. kernelspace vs userspace Kernel: ✗ T oo slow for high-speed forwarding ✗ Missing optimizations (batching, caching, etc) Userspace: ✗ No direct access to NIC (context switch, sk_bufg) ✔ DPDK (DMA, I/O batching) ✔ Software optimizations ✔ Flexibility 7 ANRW 2019

  8. Kernel-Bypass Frameworks 8 ANRW 2019

  9. architecture Vector Packet Processing (VPP) ● DPDK ● RSS queues, Zero-Copy and more ● Packet vectors ● Modular node-based processing ● Low-level optimizations (caching, pipelining) ANRW 2019

  10. architecture VPP Dual-Loop while (n_left_from >= 2) { /* prefetch next iteration */ if (PREDICT_TRUE(n_left_from >= 4)){ vlib_prefetch_bufger_header(b[2], STORE); vlib_prefetch_bufger_header(b[3], STORE); } process(b[0]); process(b[1]); b += 2; next += 2; n_left_from -= 2; } /* process remaining packets */ while(n_left_from > 0){ process(b[0]); b += 1; next += 1; n_left_from -= 1; } 7/22/19 10 ANRW 2019

  11. architecture mmb: A VPP middlebox Goals: ● Various middlebox policies (fjrewall, NAT, traffjc engineering) ● Fast even with thousands rules ● Intuitive CLI 11 ANRW 2019

  12. architecture mmb: CLI grammar 7/22/19 12 ANRW 2019

  13. architecture mmb: forwarding graph 7/22/19 13 ANRW 2019

  14. architecture mmb: forwarding graph ● Classifjcation: (Packet & Mask) ⊕ Key ● Rewrite (Packet & Mask) | Key 7/22/19 14 ANRW 2019

  15. architecture mmb: processing path 7/22/19 15 ANRW 2019

  16. measurement Performance Analysis ● FastClick: – Fast (multi-queue, ZC forwarding, batching, DPDK) – Click ● eXpress Data Path (XDP): – In-Kernel – eBPF ● iptables ANRW 2019

  17. measurement Performance Analysis: Testbed Direct Indirect PCI Passthrough Bridged 7/22/19 17 ANRW 2019

  18. measurement Performance Analysis: Baselines ● VPP, FastClick, 4.15 > 99% of direct baseline 7/22/19 18 ANRW 2019

  19. measurement Performance Analysis: 5-tuples fjrewall ● Stateless matching on 5-tuples (saddr, daddr, sport, dport, proto) 7/22/19 19 ANRW 2019

  20. measurement Performance Analysis: 5-tuples fjrewall ● Stateless matching on 5-tuples (saddr, daddr, sport, dport, proto) ● mmb & XDP at direct baseline ● FastClick matching (IPFilter) has performance issues ● Iptables 4.15 sustains direct baseline with up to 1,000 rules 7/22/19 20 ANRW 2019

  21. measurement Performance Analysis: stateful fmow matching ● Stateful matching on 5-tuples (saddr, daddr, sport, dport, proto) 7/22/19 21 ANRW 2019

  22. measurement Performance Analysis: stateful fmow matching ● Stateful matching on 5-tuples (saddr, daddr, sport, dport, proto) ● mmb & XDP at direct baseline ● FastClick at 85% direct baseline ● Iptables stateful is similar to stateless (with few rules). 7/22/19 22 ANRW 2019

  23. measurement Performance Analysis: TCP Options ● Matching on TCP Options ● Not applicable to iptables, FastClick & XDP ● Stable until 78 rules 7/22/19 23 ANRW 2019

  24. measurement Conclusion & Next steps ● mmb sustains line rate for difgerent use cases ● Next Step: Payload reconstruction ● https://github.com/mami-project/vpp-mb 24 ANRW 2019

  25. measurement Thanks ! 7/22/19 25 ANRW 2019

  26. measurement Performance Analysis: Testbed ● Intel Xeon E5-2620 2.1GHz, 16 Threads, 32GB RAM ● Intel XL710 2x40GB NICs ● Huawei CE6800 switch ● Debian 9 26 ANRW 2019

  27. measurement Performance Analysis: RTT 7/22/19 27 ANRW 2019

  28. measurement Performance Analysis: CPU time 7/22/19 28 ANRW 2019

Recommend


More recommend