Section: New Results

Using heterogeneous memories

Heterogeneous memories, such as the MCDRAM in the Xeon Phi architecture, with different latency and bandwidth characteristics, complexify the way the users allocate and use memory. In 2017, we have designed, in collaboration with CEA, an automatic tool to characterize the bandwidth needs of an application, in particular finding the functions and the arrays in these functions that would benefit the most of a high bandwidth. This tool is a plugin of gcc, and has been applied successfully to large CORAL benchmarks (Lulesh, MiniFE, AMG2013, Mcb and Snap). This characterization is essential in the common case where all data cannot fit into the MCDRAM but a more selective use of the MCDRAM is needed. The transformation of the memory allocations is automatic, based on these metrics. The development of better metrics, allowing to choose the most appropriate array is on going work.