Special Report: Network Provisioning
Statistical Inference and Optimized Data Collection: Due to the sheer data volumes, it is inefficient to collect measurements from all nodes all the time for the purposes of diagnosis, optimization and performance tuning. The measurement and data collected must be aided by systematic inferencing methods to identify the critical and canonical sets of measurements needed. Statistical methods such as involved in the design of experiments, are required to ensure that the measurements are strategic and optimal.
6.2. Integration and Interactions The network functionalities must be made available to the end users in a transparent manner independent of the underlying provisioning modes and the associated transport methods. Such levels of transparency can only be achieved by focused integration efforts, which have been often ignored by technology developers. As result, the application performance could be suboptimal and solutions could be hard to use. Both provisioning and transport methods must be developed to gracefully interface and integrate with each other as well as with the other components such as the operating system, middleware, and applications. In particular, to realize the require throughputs between the applications at the channel end-points, it is essential that the "impedance matches" be achieved at all the interfaces. Furthermore, these technologies must smoothly interact and co-exist with legacy networks, and provide a smooth transition to newer networks in the appropriate cases.
Middleware and Application Interfaces: In IP networks it is common for the applications to interface with the transport modules which in turn communicate with the lower level services. The situation of on-demand provisioning is somewhat different. The applications themselves can request and be granted the dedicated channels to be used exclusively. Both the middleware and applications must be provided the needed interfaces to the provisioning and transport modules. Indeed, it would be most desirable if the transport and provisioning methods are developed in close association with the middleware, applications and end users. Test-beds outlined in the next section could facilitate such activities.
Hardware and Operating Systems: Due to the sheer speeds of the network connections needed in large-scale applications, the usual host controlled transport methods may not be adequate. For example, a low-end host might not be fast enough to produce and/or consume the data that is arriving at multiple tens of Gbps. The network interfaces and the end hosts must be designed to operate at network speeds. Special modules may be necessary to implement OS-bypass and RDMA methods to relieve the host CPUs from being exclusively consumed by data transfers. Also host clusters might be required to operate in striped mode to sustain data rates that are commensurate with the channels and end-systems such as High Performance Storage Systems (HPSS).
Previous Next Table
of Content for report: Network Provisioning Home
