We give a brief overview of our work on Invasive Computing within the “Invasive Computing”1 transregio (TR) project which inspired the application of the invasive paradigm also in HPC.
Thousands and more heterogeneous cores are expected on future multi-processor systems on chips [Tei08]. For concurrently executed applications or applications with changing resource requirements during runtime, this demands an orchestration of resources. Our approach is based on giving the application the possibility to specify resource requirements during run time with the invasive commands invade, infect and retreat described in the previous section. E.g. if an application fails to yield a particular parallel efficiency due to reduced workload, it can retreat from resources. This makes the resources available to other applications. To support these invasive commands in an efficient way, several challenges have to be solved such as scheduling of resource-competing applications, programmability, algorithmic redesign, new hardware components, etc. Contrary to our HPC approach which relies on extensions to standard parallelization models and software extensions only, the invasive hardware platform developed in the TR also offers invadable hardware components on a multiprocessor systems-on-a-chip (MPSoCs) and an invasive software tool chain.
Our main task in this TR is to provide the knowledge on algorithms from the area of scientific computing. Based on the requirements of our algorithms, we took part in the decision making of several developments of the Invasive System on the hard- as well as software layers. To show the differences to our HPC approach, we continue to give a brief introduction to the Invasive System.