DPDK Userspace 2018

We would like to present the requirement and concerns of hyper scale users for support of NFV's and virtual switches/routers over DPDK which require HW to support offload of millions of rules, connection tracking and traffic counters, as well as hundreds of thousands of packet modification actions.

Multiple layers of rule caching, in rte_flow, PMD, application copy and in HW makes a huge memory footprint for applications. Supporting batch processing instead of single entry processing, like for reading of counters, multi-rule deletion/update is missing for allowing high rate rule updates.

Enabling async API's can release the application CPU while HW and firmware are updating the smart NIC offloads and allow even more.