Figure 3: A scenario for fusion of Isosurface calculation front-end
and back-end.
Agent fusion is applied to those components in the sample applications for which co-location on the same machine and with the same internal form (native or neutral) is likely to occur. For instance, Isosurface calculations in the ISDA application may be performed outside the scope of the visualization engine (if the visualization runs on a remote or weakly connected machine) and jointly with visualization. In fact, earlier versions of ISDA always performed Isosurface calculation within the visualization agent itself, as the visualization was running on a strongly connected and high end visualization engine (i.e., an SGI Octane). The versions of ISDA now used by end users do not wish to use such an instance of the visualization agent due to their desire to operate across a wider spectrum of machines and service a larger number of end users.
Figure 3 depicts the situation in which fusion is carried out on the Isosurface front-end and back-end, where the back-end does the actual isosurface calculation and then sends the computed isosurfaces as data to front-end on which visualization agent runs. Originally, there are multiple Isosurface front-ends, and the back-end is not co-located with any of the front-ends, with the intent of minimizing overall communication cost and avoiding burden the visualization engines with isosurface calculations. When all front-ends except one complete their execution, then the back-end might migrate to the remaining front-end's location, in order to reduce the communication cost between the two parties. The system is then able to fuse the two agents to further enhance performance. Table 4 shows the potential performance improvements resulting from such a fusion action. The same table also lists the results of fusing two agents in the PSSPS application: fir filtering and range processing.
In both of the experiments shown below, fusion is done manually through simulating the actions would have been taken by the fusion compiler. Such actions include procedure in-lining, data sharing, and asynchrony elimination.
Table 4: Result of fusion applied to different agent
forms.