Download presentation
Presentation is loading. Please wait.
1
Flurries: Countless Fine-Grained NFs for Flexiable per-flow Customization
Wei Zhang, Jinho Hwang, Shriram Rajagopal, k.k. ramakrishnan and timothy wood – conext 2016
2
From: Class Wide Monolithic NFs To: Light-Weight Flow Specific NFs
Dedicated Core NF's on containers share other cores 0-copy by shared memory – no full isolation DPDK: NIC <-> RX\TX RX --> Flow Table --> NF #i Rx RX --> Flow Director --> Controller --> allocate NF Costly statistics gathering for fine-grained scheduling.
3
Per-Flow NF's & NF's state transation
NF's life-time is as long as the flow is active. NF's are woke up By Wakeup thread and not by every packet arrival (threshold / timeout). When flow finishes, the NF resets its state and is moved to the idle pool. NF's containers should use the libflurries API. No dynamic idle pool resizing feature.
4
80,000/s Flurry NFs, 30 Gbps 28 core Intel Xeon 2.6 GHz. 3 Dual Port NIC’s (10G). 164 GB Memory.
5
Chain-aware scheduling
Overloaded system. Chain 1:3:7. Drop as early as possible in the chain. Unaware => same priority => NF1 hogs the cpu Flurries => later NF’s have higher priority
Similar presentations
© 2024 SlidePlayer.com Inc.
All rights reserved.