Using Sharable Mempools for Zero-copy Sharing Between Processes - Bruce Richardson, Intel

Using Sharable Mempools for Zero-copy Sharing Between Processes - Bruce Richardson, Intel

This talk focuses on the issues associated with having multiple DPDK processes which wish to co-operate together in a high-performant manner. For example, within a single workload, it may be desirable to split up the work across multiple independent processes rather than separate threads. In such cases, the standard IPC mechanisms for passing packets between these independent processes may be suboptimal - generally involving at least one, if not two - data copies, as well as descriptor manipulation e.g. with virtio or memif. This can lead to the scenario where the developer may decide to use DPDK multi-process mode, instead of having fully independent processes. This gives very good performance and efficient buffer transfer between processes - but at a cost of flexibility and resilience. This talk therefore examines the problem from a different angle. It discusses ways in which we can create special sharable blocks of memory for use by multiple processes, and the gory details of how it may be possible to create memory pools that can be shared by processes without requiring identical memory mappings between processes.

Accelerating Network Performance - The most popular kit to accelerate packet processing on a wide variety of CPU architectures Designed to run on x86, POWER and ARM processors, DPDK is a set of libraries running mostly in Linux userland, with a FreeBSD port available for a subset of features. https://www.dpdk.org/