InfiniBand is a new method of communication between I/O devices in the internet infrastructure. It's a switch fabric communication link which supports quality of service and fail overs providing a highly scalable solution for high performance computing with high availability.
Just like in other existing common communication channels, Fiber channel, PCI express, Serial ATA InfiniBand supports point to point communication, especially connnceting servers and storage disks. Diagram below shows how the InfiniBand comes in to play with a software implementation.
When comparing the features, InfiniBand have a higher bandwidth compared to other solutions. But Hyper transport channel has a higher max bandwidth allocation. As the transport medium, all three methods PCB, Copper and Fiber exist providing much more flexible solution at the same time providing much more signal length in both the fields, PCB and Copper. Another highlighting feature of InfiniBand is the maximum packet payload, 4KB which is very large relative to the values of the other channels.
InfiniBand Layers
The architecture of InfiniBand is divided into many layers which operates independant of one another. Mainly there are four divisions, Physical, Network, Transport and Upper layers.
Physical layer specifies the actual hardware, cables, backbones and tangible equipment which needs in InfiniBand communication.
Link layer is the heart of the architectur which handles three core tasks. Packet layout definition, point to point link operations and switching between local sub net.
Network layer handles routing paths between sub nets. It makes sure the packets are delivered from one subnet to another subnet.
Transport layer is placed in order to verify guranteed delivery. Partitioning, channel multiplexing and transport services are the services provided in transport layer.
InfiniBand Architecture
No comments:
Post a Comment