Networking technologies offered by exhibitors at SuperComputing '94 included, but were not limited to:
All networking products discussed in this paper, claim to have some LAN utility --- even those networking technologies that are more closely associated with point-to-point communications. These capabilities are due in part to a common paradigm used in the protocol stack that controls how applications access the networking hardware. Vendors for SCI, HiPPI, Fibre Channel, and ATM technologies have all implemented a software protocol access paradigm that provides two distinct ways for an applications to access a network interface. One way for an application to access the network interface is to include direct access to the UNIX device drivers for the interface through an application programming interface (API). The other part of the interface access paradigm provides transparent integration of the networking technology to existing applications through the Internet protocol (IP) socket-based protocol stack, and provides transparent, but sometimes crude, LAN capabilities. Figure 4 illustrates this two option paradigm. The most significant aspect of this dual protocol paradigm is that existing applications can utilize the networking technology without having to undergo costly modifications to the software. Utilizing the transparent access to the interface through the IP stack causes additional overhead when compared to the API that directly accesses the device drivers. There is additional overhead involved in implementing the transmission control protocol (TCP)/IP protocol, especially if it is not required because lower layer protocols ensure data reliability.
Figure 4: Dual Protocol Paradigm for Interfacing Applications to Network Hardware
Multi-computer clustering is an important application that uses these various networking technologies to interconnect multiple computers into a (poor-man's) scalable parallel processor. Considering the obvious fact that as interprocessor communications latency decreases and bandwidth increases, more classes of applications and algorithms can be efficiently implemented on parallel processors --- those technologies with software protocols and hardware designs that provide low latency and ample bandwidth will provide the most robust clustered computing environments. Latency and bandwidth are important to computer scientists designing algorithms and software engineers porting applications; however, system cost is also a factor that must be considered.
Throughout the remainder of this section, I will describe these six networking technologies considering