12.1. Understanding InfiniBand and RDMA technologies
InfiniBand refers to two distinctly different things. The first is a physical link-layer protocol for InfiniBand networks. The second is a higher level programming API called the InfiniBand Verbs API. The InfiniBand Verbs API is an implementation of a remote direct memory access
RDMA provides direct access from the memory of one computer to the memory of another without involving either computer’s operating system. This technology enables high-throughput, low-latency networking with low CPU utilization, which is especially useful in massively parallel computer clusters.
In a typical
IP data transfer, application X on machine A sends some data to application Y on machine B. As part of the transfer, the kernel on machine B must first receive the data, decode the packet headers, determine that the data belongs to application Y, wake up application Y, wait for application Y to perform a read syscall into the kernel, then it must manually copy the data from the kernel's own internal memory space into the buffer provided by application Y. This process means that most network traffic must be copied across the system's main memory bus at least twice (once when the host adapter uses DMA to put the data into the kernel-provided memory buffer, and again when the kernel moves the data to the application's memory buffer) and it also means the computer must execute a number of context switches to switch between kernel context and application Y context. Both of these things impose extremely high CPU loads on the system when network traffic is flowing at very high rates and can make other tasks to slow down.
RDMA communications differ from normal
IP communications because they bypass kernel intervention in the communication process, and in the process greatly reduce the CPU overhead normally needed to process network communications. The RDMA protocol allows the host adapter in the machine to know when a packet comes in from the network, which application should receive that packet, and where in the application's memory space it should go. Instead of sending the packet to the kernel to be processed and then copied into the user application's memory, it places the contents of the packet directly in the application's buffer without any further intervention necessary. However, it cannot be accomplished using the standard Berkeley Sockets API that most
IP networking applications are built upon, so it must provide its own API, the InfiniBand Verbs API, and applications must be ported to this API before they can use RDMA technology directly.
Red Hat Enterprise Linux 7 supports both the InfiniBand hardware and the InfiniBand Verbs API. In addition, there are two additional supported technologies that allow the InfiniBand Verbs API to be utilized on non-InfiniBand hardware:
The Internet Wide Area RDMA Protocol (iWARP)
iWARP is a computer networking protocol that implements remote direct memory access (RDMA) for efficient data transfer over Internet Protocol (IP) networks.
The RDMA over Converged Ethernet (RoCE) protocol, which later renamed to InfiniBand over Ethernet (IBoE).
RoCE is a network protocol that allows remote direct memory access (RDMA) over an Ethernet network.
Both iWARP and RoCE technologies have a normal
network link layer as their underlying technology, and so the majority of their configuration is actually covered in Chapter 3, Configuring IP Networking
. For the most part, once their
networking features are properly configured, their RDMA features are all automatic and will show up as long as the proper drivers for the hardware are installed. The kernel drivers are always included with each kernel Red Hat provides, however the user-space drivers must be installed manually if the InfiniBand package group was not selected at machine install time.
Since Red Hat Enterprise Linux 7.4, all RDMA user-space drivers are merged into the rdma-core
package. To install all supported iWARP, RoCE or InfiniBand user-space drivers, enter as
yum install libibverbs
If you are using Priority Flow Control (PFC) and mlx4-based cards, then edit
/etc/modprobe.d/mlx4.conf to instruct the driver which packet priority is configured for the “no-drop” service on the Ethernet switches the cards are plugged into and rebuild the
initramfs to include the modified file. Newer mlx5-based cards auto-negotiate PFC settings with the switch and do not need any module option to inform them of the “no-drop” priority or priorities.
With these driver packages installed (in addition to the normal RDMA packages typically installed for any InfiniBand installation), a user should be able to utilize most of the normal RDMA applications to test and see RDMA protocol communication taking place on their adapters. However, not all of the programs included in Red Hat Enterprise Linux 7 properly support iWARP or RoCE/IBoE devices. This is because the connection establishment protocol on iWARP in particular is different than it is on real InfiniBand link-layer connections. If the program in question uses the librdmacm connection management library, it handles the differences between iWARP and InfiniBand silently and the program should work. If the application tries to do its own connection management, then it must specifically support iWARP or else it does not work.