New Publication: H2M – Exploiting Heterogeneous Shared Memory Architectures

In the DFG-funded project "Heuristics for Heterogeneous Memory" (H2M), RWTH Aachen University and the French project partner Inria are jointly developing support for new memory technologies such as High Bandwidth Memory (HBM) and Non-Volatile Memory (NVRAM). These technologies are increasingly being used alongside traditional Dynamic Random Access Memory (DRAM) in HPC systems. HBM offers higher bandwidth but smaller size than DRAM. NVM offers greater capacity but is slower than DRAM. Given these differences, the question is how to efficiently use systems with heterogeneous memory and where data should be stored.

Currently, using new memory technologies requires significant modifications to applications and the use of platform- or vendor-specific application programming interfaces (APIs). The goal of H2M is to provide portable interfaces to identify and provide access to available storage. On this basis, allocation abstractions and heuristics are developed to give application developers and runtime systems control over where data should be stored and when data should be moved between different types of storage.

The usual memory allocation functions like malloc (in C) and new (in C++) are abstracted by a new allocation function (h2m_alloc_w_traits). Application developers can pass additional properties (traits) to this allocation, which describe, for example, how the data is to be used, accessed or which requirements the allocation must fulfil.

The H2M runtime system can then determine the appropriate memory to store the data using strategies that consider both traits and the types of memory available in the system. The new concept will be evaluated with various proxy applications, so-called mini-apps, and benchmarks on different architectures. It will also be investigated whether dynamic movement of data between storage types can lead to efficiency gains, especially in scenarios where data objects are used differently in different execution phases or access patterns to data change significantly.

The full publication can be found at ScienceDirekt.

For more information, see the project website.