About Infiniband

infinibandA high-throughput I/O technology, Infiniband is based on a switched fabric of serial data streams. Attaining bandwidths of 2.5 to 30Gbs, it exceeds the fault tolerance and scalability limits of shared bus architecture “through the use of switches and routers in the construction of its fabric” (Pentakalos, 2002). Starting out as two separate initiatives, Future I/O founded by Compaq, IBM, and HP, and the Next-Generation I/O Initiative by Dell, Hitachi, Intel, NEC, Siemens, and Sun, a single entity was eventually formed: the Infiniband Trade Association (ITA). The main reason that Infininband Architecture (IBA) was developed was because processing power was quickly “outstripping the capabilities of industry-standard I/O subsystems using busses” (Pfister, undated). Part of the speed within this architecture is achieved through a serial connection: four connections as opposed to wide parallel on a PCI bus. Infiniband also provides, within the switched fabric (which provides multiple paths to a storage target), the capability of sharing storage targets among multiple servers and the “ability to perform third party I/O” (Pentakalos, 2002). This means that the storage device(s) can complete the I/O transaction without the contribution of a host: freeing CPU cycles on the host and speeding up application response times. Basically, Infiniband is a point-to-point connection, not bussed. This provides fault isolation and avoids arbitration within data flow, and can scale to a large extent because of a switched network element. Implemented within high performance computing (HPC) clusters, infiniband provides “high bandwidth and low message latency attributes to inter-processor communication systems” (Lugones, et al, 2008).

Reference:

Pentakalos, O. (2002). An Introduction to the Infiniband Architecture. Windows 2000 Performance Guide. Oreilly. Retrieved August 14, 2009 from http://www.oreillynet.com/pub/a/network/2002/02/04/windows.html

Pfister, G. (undated). An Introduction to the Infiniband Architecture. IBM Enterprise Server Group Server Technology and Architecture. p.617.

Lugones, D., Franco, D., Luque, E. (2008). Dynamic routing balancing on InfiniBand networks. Journal of Computer Science & Technology 8.2 Academic OneFile. Gale. BCR Regis University. 14 Aug. 2009

Advertisements