CaDS Seminar 2022 - May 24

Godehard Sutmann (SDL Molecular Systems)

Load Balancing Revisited: From Average to Maximum

Abstract:
The properties of tensor decomposition load balancing method are analysed and discussed in terms of quality of the work distribution. It is considered, which properties work load functions need to be distributed ideally and which tolerance needs to be considered for more general distribution functions. We consider an optimisation of the tensor decomposition via least squares and introduce a new and efficient decomposition method, based on the minimisation of work load maxima, which can be derived from the usual tensor decomposition method. Superior behaviour of the new method is demonstrated and analysed in terms of an Eigenvalue decomposition of the underlying iteration matrix. Similarities to Markov chain transitions are discussed.