Grid computing has been defined as "coordinated resource sharing and problem solving in dynamic, multi-institutional virtual organisations" - in practice, this usually means that a computation intensive application is distributed across a number of computers for the sake of speeding it up. This step from local clusters to massively distributed applications that use the Internet as its underlying means of communication entails a requirement for efficient network usage, but the Internet is a general-purpose network that was not built to provide maximum efficiency to the Grid.
EC-GIN has properties that make it unique from a networking perspective, and different from classical networking; key examples are listed below.
a) A number of nodes at different locations contribute to a common cause, and they can be expected to remain available for a long time; which differs from a typical client/service model, where a node at a single location traditionally remains available.
b) Traffic typically consists of short, sporadic method invocations (Grid Service calls) and bulk data transfers, which may be huge. The distinction between these two types of data transfer is usually clear to the programmer of a Grid application, who explicitly invokes a file transfer (e.g., by calling the "GridFTP" service) when she or he sends a large amount of data.
c) File transfers are "pushed" and not "pulled": in a standard FTP session, it is more common for users to initiate file downloads rather than uploads. In a Grid scenario, a scheduler decides where application parts go; then, it transfers the required data to all the places where calculations are carried out. This means that there may be more than one receiver for a large file, which renders reliable multicast especially useful in a Grid setting.
d) A Grid application may be able to specify its communication processes in advance. In general, Grid applications have a larger probability of showing some workflow aspects than more traditional applications. Thus it is reasonable to have a networking-related analysis performed automatically; this can be based on workflow information and other heuristics.
e) Typically, Internet Service Providers do not need to align their networks to applications, however, for the Grid traffic type expected and while optimising network services in general, suitable and efficient accounting mechanisms in support of a decentralised Grid are viable. Incentive schemes will provide for Grid application centric approaches which will adapt to the existing physical network structures.
f) Relevant Grid service management approaches will be required to be supported by appropriate security and trust mechanisms, which address in an efficient and viable manner the distributed Grid services environment.