Small Tree's InfiniBand technology
Oct 1, 2005 12:00 PM, By Corky Seeber
In the broadcast industry, content delivery bandwidth is critical, and today's network wiring typically offers no more than a gigabyte of throughput. As the industry moves toward HD, a network-wiring alternative is needed to handle increased data requirements. The emerging solution is InfiniBand — a one-wire, high-performance interconnect supported by Small Tree Communications.
Moving data quickly
Traditional networking protocols are directed through a TCP/IP stack. TCP was designed more than 20 years ago, at a time when files were smaller and networks were slower and less reliable. Packet sizes were limited to 1500B as a tradeoff between using large packets for efficiency, while keeping them small enough so a small amount of corruption would not require too much data to be retransmitted. Each of these tiny packets is encapsulated with a medium access control (MAC) header, an IP header and a TCP (or perhaps UDP) header, which all have to be stripped off at the final destination. This leads to quite a bit of overhead and protocol traffic, creating additional latency and congestion and making a traditional TCP/IP network less than ideal for editing HD footage in real time.
The idea behind the InfiniBand technology was to create a scalable, extensible fabric that could be used to interconnect systems, I/O devices and storage. The latency to send a message across the fabric needed to be extremely low in order for things like MPI clustering and storage to be effective.
The technology's low-level protocols allow upper level protocols to be set on top of them in such a way as to add minimum overhead. Sockets Direct Protocol (Expected Q106) is one such example. It allows sockets-based applications to send data directly through the InfiniBand network without using TCP. As a result, there is much less protocol overhead, and updating frames can happen more quickly. Latencies on the TCP stack on Mac have been measured at 60µs, even with 10Gb, whereas when using InfiniBand, latencies have been measured at 6µs.
So just what is this one-wire technology? It's a switch-based serial I/O interconnect architecture that operates at a base speed of 20Gb/s per port, or 10Gb/s in each direction. It differs from shared bus architectures in that it is a low pin-count serial architecture that connects devices on the PCB and enables bandwidth out of the box. It provides both the high bandwidth one would expect from a 10Gb/s interconnect as well as the low latency advantages one might see when communicating with a device over a local bus. The architecture was designed to simplify and speed server-to-server connections and links to additional server-related systems.
The primary use of this technology is focused in cluster systems — two or more systems working as one. It was originally created to serve as a bus replacement, acting as a switched fabric network in place of 10Gb or other Ethernet networks. Today, more and more broadcast companies are looking carefully at cluster file system technologies. These technologies are viewed as a way of storing all data in one place and helping reduce copies.
A problem with current cluster file systems is latency to and from the metadata server. Somewhere in the organization, there is a server tracking who is working with a file. It also prevents two people from working from the same file at the same time. Unfortunately, almost every one of these metadata servers cannot provide greater than 40µs to 60µs latency from the time the message is sent out and the other side receives it. This technology offers the promise of reducing latency by a full order of magnitude to improve the ability of clustered file systems to operate across a group of machines at greater speeds.
Shifting to one wire
As a one-wire network interconnect, the InfiniBand technology also helps reduce infrastructure requirements. Current workstation systems easily have four cables — one or two each for Ethernet and Fibre Channel — plus a disc drive connected to one system.
With Small Tree's technology, each individual workstation in a workgroup of 20 users, for example, would run their workstations via one wire to a relatively small switch that is bridged back to a larger switch in the data center. This gives each user up to 10Gb/s of access to all the data stored on the server. Using bridged I/O, none of these systems need to have Ethernet, and you can get rid of the discs within those systems.
Corky Seeber is president of Small Tree Communications.
| Want to use this article? Click here for options! |  |
