Summary of EN141 Webinar: RoCE Introduction

00:00:00 - 00:15:00

The video introduces the concept of remote direct memory access (RDMA) and describes how it works. It also covers the benefits of RDMA over traditional methods such as the TCP/IP stack. Finally, the presenter covers different RDMA setups and how to signal an RDMA transfer to start.

• 00:00:00 In this video, the presenter introduces the concepts of remote direct memory access (RDMA) and Rocky solution, and covers the benefits of RDMA over traditional methods such as the TCP/IP stack. They then introduce some basic RDMA concepts and describe how an RDMA transfer takes place. Finally, the presenter covers different RDMA setups and how to signal an RDMA transfer to start.
• 00:05:00 RDMA is a technology that allows for the efficient transfer of large amounts of data between systems. The technology is well suited for applications that require the transport of large amounts of data, such as artificial intelligence and machine learning applications, as well as storage apps, such as VMware vSAN. The RDMA verbs layer provides an interface for the operation of RDMA applications, and the RDMA transport and network layers are implemented inside the neck of the hardware. The RDMA concepts covered so far are independent of the underlying network fabric.
• 00:10:00 Broadcom's "Rocky" control path and data path operating system is supported on Windows and VMware ESX. The software and hardware components fit into the Linux RDMA stack, with the Rocky user space library and kernel space driver communicating with the Nic firmware and neck Hardware over PCie. The data path is completely implemented in the Nik Hardware, and high performance Computing HPC applications make extensive use of RDMA. Many HPC applications make use of the Message Passing Interface MPI, which is a specification for parallel Computing architectures. Broadcom's next support the prmm direct feature, which allows RDMA across peer devices in this context. Prmm direct allows RDMA across two peer device memories, where the two devices can be on two different hosts on the network. For prmm-based applications, if the RDMA happens across host memories only the CPU still has to transfer the data from the host memory to the local peer device memory. PRMim direct allows two remote devices to directly RDMA across each other's memory via Broadcom Lake. This transfer of data from the Nick to the peer device memory and vice versa happens over the pcie bus using pcie peer-to-peer transfers. For peer mem direct, an additional kernel module, known as
• 00:15:00 This webinar introduces Broadcom's RoCE congestion control solution, which is used in conjunction with the company's next-generation networking platform, Broadcom Nix. The webinar covers the basics of RoCE congestion control, including how the solution works and how to configure it.