nsdi2015-slides
nsdi2015-slides nsdi2015-slides
Implications for Forwarding Performance OpenFlow tables Table 0 Table 1 packet Flow 1 Flow 1 ingress ... Flow 2 Flow 2 ... ... Table 24 Flow 1 Flow 2 ... packet egress Physical to Logical L2 Lookup ... Logical to Physical k 0 hash lookups k 1 hash lookups ... k 24 hash lookups 100+ hash lookups per packet for tuple space search?
Non-solutions ● All of these helped: – Multithreading – Userspace RCU – Batching packet processing – Classifier optimizations – Microoptimizations ● None of it helped enough: % versus x. Classification is expensive on general-purpose CPUs!
- Page 1 and 2: The Design and Implementation of Op
- Page 3 and 4: Where is Open vSwitch Used? ● Bro
- Page 5: Use Case: Network Virtualization Op
- Page 9 and 10: Microflow cache (1 hash lookup) pac
- Page 11 and 12: Using a More Expensive Cache OpenFl
- Page 13 and 14: Lazy Approach to Populating Cache S
- Page 15 and 16: Making Megaflows Better ● Megaflo
- Page 17 and 18: OVS Cache v3: Dual Caches μ hit Mi
Non-solutions<br />
●<br />
All of these helped:<br />
– Multithreading<br />
– Userspace RCU<br />
– Batching packet processing<br />
– Classifier optimizations<br />
– Microoptimizations<br />
● None of it helped enough: % versus x.<br />
Classification is expensive on general-purpose CPUs!