GRO Library Enhancements - Kumara Parameshwaran Rathinavel, Microsoft

GRO Library Enhancements - Kumara Parameshwaran Rathinavel, Microsoft
The current GRO library in DPDK is suboptimal. For every packet, to verify the presence of the current 5-tuple in the GRO table, a lookup is performed on each flow. Implementing a hash-based solution for the 5-tuple would be an efficient optimization. Furthermore, it would be advantageous if applications could configure the timeout for a specific flow or 5-tuple. The existing timer mode in GRO applies a static timeout for any application, which is not ideal. The timeout should be adjustable based on the latency sensitivity of the application. If the GRO layer processes flows from different applications, a single timeout setting could lead to suboptimal performance for others. Providing an infrastructure within GRO for applications to set timeouts for tuples would be beneficial. This would allow for the judicious use of GRO resources, optimizing for different types of traffic, such as mice versus elephant flows. If packets from multiple applications are processed by the GRO layer, the varying latency tolerances could be accommodated by setting appropriate GRO timeouts.

[@darthhame]

Very insightful & innovative talk!