This section deals with FAST, a performance prediction module that can
be used by DIET. It is non-mandatory, but can provide SeDs with
improved performance prediction capability.
You can use FAST in stand-alone mode without having compiled with
CoRI option.
FAST [21] is a tool for dynamic performance forecasting in a Grid environment. As shown in Figure 8.1, FAST is composed of several layers and relies on a variety of low-level software. First, it uses the Network Weather Service (NWS) [25], a distributed system that periodically monitors and dynamically forecasts the performances of various network and computational resources. The resource availabilities acquisition module of FAST uses and enhances NWS. Indeed, if there is no direct NWS monitoring between two machines, FAST automatically searches for the shortest path between them in the graph of monitored links. It estimates the bandwidth as the minimum of those in the path and the latency as the sum of those measured. This allows the availability of more predictions when DIET is deployed over a hierarchical network.
In addition to system availabilities, FAST can also forecast the time and space needs of certain computational routines as a function of the problem parameters and the machines where the computations would take place. FAST is particularly suited to numerical algebra routines whose performance is not data-dependent and where a clear relationship exists between problem size and performance. As a basis for predictions, FAST benchmarks the routines at installation time on each machine for a representative set of parameters. After polynomial data fitting, the results are stored in an LDAP tree. The user API of FAST is composed of a small set of functions that combine resource availabilities and routine needs from low-level software to produce ready-to-use values. These results can be combined into analytical models by the parallel extension [3] to forecast execution times of parallel routines.
FAST clients can access information like the time needed to move a
given amount of data between two FAST-enabled machines SeDs, the
time to solve a problem with a given set of computational resources,
or the combination of these two quantities.
For more details about FAST, please refer to the FAST webpage 7.1.
