The International Conference for High Performance Computing, Networking, Storage and Analysis
Analysis of Energy and Power Consumption by Remote Data Transfers in Distributed Programming Models.
Student: Siddhartha Jana (University of Houston)
Supervisor: Barbara Chapman (University of Houston)
Abstract: Recent studies of the challenges facing the Extremescale era express a need for understanding the impact on energy profile of applications due to inter-process communication on large-scale systems. Programming models like MPI provide the user with explicit interfaces to initiate data-transfers among distributed processes.
This poster presents empirical evidence that: controllable factors like the size of the data-payload to be transferred and the number of explicit calls used to service such transfers have a direct impact on the power signatures of communication kernels. Moreover, the choice of the transport layer (along with the associated interconnect) and the design of the inter-process communication protocol are responsible for these signatures. Results discussed in this poster motivate the incorporation of energy-based metrics for fine tuning communication libraries that target Extremescale machines.