Crying for InfiniBand, InfiniBand
Since ancient times, there have been some wonderful things that have not ended before. I call them martyrs.
InfiniBand is one of them. Although it has fallen, I still want to applaud it.
If the purpose of Ethernet is to link hosts together, InfiniBand's original intention is to dismember them. The difference between these genes is doomed to a war.
The server is powerful, but the PCI bus is too slow, and its parallel design is not suitable for high-speed environments at all. The correct method is serialization and clock frequency increase, because parallelism cannot handle clock frequency improvement, it will crash. InfiniBand is designed to pull the bus out of the chassis and provide a high-speed serial transmission path. It is successful because it separates the end-to-end complexity and is no longer determined by the bus itself, this idea is correct, just as the evolution from the CSMA/CD Broadcast Network to the switched Ethernet network is the same.
The control logic is migrated from the link itself to the two ends of the link. So the hub came out. For a hub without a controller, as the Chinese name says, it is a hub. Then the concept of a bridge is introduced, and the Ethernet bridge becomes a switch, so far, InfiniBand is also taking the same route. It may realize that parallel transmission and high-frequency clock cannot have both sides, so it uses high-frequency Clock and serial mode, just like switching Ethernet, the bus topology is changed to a point-to-point topology and uses message routing, which is exactly the same as the Ethernet frame. If InfiniBand continues to this path, not only replaces PCI, PCI-X, but also is likely to compete with Ethernet, Ethernet from the birth of no defeat, are relying on cost KO opponent, although InfiniBand is a little short in this regard, it has enough time to improve. Didn't Ethernet go through the expensive coaxial age at first?
The InfiniBand and Ethernet architectures are really similar, but one difference is that InfiniBand is flat, while Ethernet is part of the network protocol stack and can interface with IP addresses, this is also his job. InfiniBand is loyal to its own position. It just wants to pull peripherals out of the chassis to form a new high-speed, scalable server bus architecture. To increase the bus bandwidth or add a new device, you just need to add a cable! This is the killer of InfiniBand. Because it can be external to the chassis, this physical separation makes the InfiniBand Switch an independent device and does not belong to the motherboard as a PCI Controller. In short, everything is perfect.
It's a bit sad to go on here...
Cost problems, coupled with the sharp drop in investment caused by the burst of the Internet bubble, InfiniBand's problems have led many of its advocates and even sponsors to leave, and both their parents and their loved ones have left. InfiniBand seems thin... however, this is not the end of the game. At this time, PCI-Express, that is, PCIE, is almost a replica of InfiniBand, and its main strength is still inside the chassis, of course, there can also be standard transfer media extended to the outside of the box. This completely makes InfiniBand a martyr. Just like Christ in Christianity.
Later on, InfiniBand was learned by many new technologies, such as using TCP/IP for remote memory access (RDMA). Of course, its Protocol Stack is fixed on dedicated hardware. Nowadays, the individual server chassis is already very crowded. The other side of ultra-fierce server virtualization is that multiple servers are used as one, while InfiniBand's thought was at that time, too early-maturing...
Tears For InfiniBand. But it is still slow, and it has not completely disappeared. It may be worth looking forward to the day when it is upgraded.
Copyright Disclaimer: This article is an original article by the blogger and cannot be reproduced without the permission of the blogger.