PEARC17

The Internet's conceptual model is that of the classic phone system: Connect two instruments (or interfaces) together with a wire. This is today an appropriate model still for some applications—e.g., connecting a remote user to the login host of an HPC system. But even this simple model admits of diversity for varied applications without analogs in classical telephony—quality of service in terms of guaranteed bandwidth and latency control, and elaborations such as the Science DMZ and DTNs.  Yet in the computational sciences, it is often the case that an investigator's focus is on the dataset(s) to be analyzed, while the interfaces and the hosts on which they reside are of no interest. These ideas lead to the concept of "information-centric networking" and network architectures different from the Internet, which are of growing interest both commercially and in the academy.  Until quite recently, network designers were forced to choose a communications model and network architecture to implement. Today however, the availability of very fast commodity hardware, inexpensive storage, and abundant bandwidth allow multiple architectures to co-exist on the same network substrate, and for particular instances to be stood up and torn down under end-user control. This talk will elaborate on these ideas and offer some examples.