Section: New Results

An interval constrained memory allocator for a GAS runtime

Participants : François Gindraud, Fabrice Rastello, Albert Cohen [ENS Ulm] .

This work presents a memory allocator for global address space (GAS) runtime targeting distributed memory embedded architectures (MPSoC). MPSoC we are interested in are relatively new architectures, composed of several nodes with multiple general purpose cores and a local memory, linked by a network, all on one chip (NoC). They have promising energy and computing performances, but are hard to program due to the multilevel parallelism and the hardware constraints (limited memory, network structure). Existing programming framework are either thin but let the programmer do the hard choices (OpenMP + MPI) or heavy and automatic but target specific kind of applications on big systems (Global Arrays).

Givy  5.1 is a runtime currently developed to execute dynamic task graphs with data-flow dependencies on MPSoC. It has a focus on supporting irregular applications, using the dependencies to perform data-aware dynamic task scheduling and data transfer. Data blocks live in a GAS, and thus requires a GAS-aware memory allocator to avoid address collisions when they are dynamically allocated. The allocator implementation proposed in this paper does this with zero synchronization between nodes, while being memory efficient in the small distributed memories, and fast on each multithreaded node.

This work will be submitted at ACM ISMM Symposium.