Despite that the Invasive Computing paradigm was originally developed for TCPAs, it showed its applicability also in areas of MPSoCs as well as HPC shared-memory systems.
For the Invasive MPSoC System with its heterogeneous computing components, we selected the multigrid solver as a representative algorithm and presented the challenges for extensions with Invasive Computing in X10. With the Invasive Computing interfaces offered by invadeX10, this considerably improves the programmability of Invasive Computing on systems with no inter-tile cache coherency.
Applying the Invasive Computing paradigm on shared-memory HPC systems comes with benefits, but also drawbacks: An obvious drawback is the fact, that applications have to provide information on their current resource requirements via the constraint system. Deriving resource requirements such as scalability graphs is challenging, however the computational workload yields an alternative. The application developer also has to consider the possibly changing number of computing resources. For applications which make use of cached thread affinities such as those shown in Sec. 8.5.2, another programming pattern has to be used. Besides these drawbacks, we proofed that Invasive Computing in HPC can lead to severe performance benefits if executing several resource-competing applications with Invasive Computing. We gained an improvement of 45% in execution time for Tsunami parameter studies compared to other parallelization models such as TBB or OpenMP.
With our large-scale simulations using dynamically adaptive mesh refinement on distributed-memory systems (see Sec. 5.12) which is e.g. required to simulate wave propagations on earth scale (see Sec. 6.5) and with multi-layers (see Sec. 6.6), we expect even more changing resource requirements for each application, hence more optimization possibilities with Invasive Computing. This demands for Invasive Computing also on HPC distributed-memory systems. Contrary to the hard- and software development for the Invasive MPSoCs, Invasive Computing on HPC large-scale systems has be based on a standard parallelization model for distributed-memory systems to allow the invasification of existing codes. However, this yields several challenges such as modifications of the MPI communicator, replication of the program instance to new computational resources and the synchronization of the simulation state. With such Invasion-enabled large-scale simulations on distributed-memory systems, we expect even more improved efficiency.