Section: New Results

Self-Stabilizing Control Infrastructure for HPC

Participants : Thomas Hérault, Camille Coti.

High performance computing platforms are becoming larger, leading to scalability and fault-tolerance issues for both applications and runtime environments (RTE) dedicated to run on such machines. After being deployed, usually following a spanning tree, a RTE needs to build its own communication infrastructure to manage and monitor the tasks of parallel applications. Previous works have demonstrated that the Binomial Graph topology (BMG) is a good candidate as a communication infrastructure for supporting scalable and fault-tolerant RTE.

In this work, we presented and analyzed a self-stabilizing algorithm to transform the underlying communication infrastructure provided by the launching service (usually a tree, due to its scalability during launch time) into a BMG, and maintain it in spite of failures. We demonstrated that this algorithm is scalable, tolerates transient failures, and adapts itself to topology changes.

The algorithms are scalable, in the sense that all process memory, number of established communication links, and size of messages are logarithmic with the number of elements in the system. The number of synchronous rounds to build the system is also logarithmic, and the number of asynchronous rounds in the worst case is square logarithmic with the number of elements in the system. Moreover, the salf-stabilizing property of the algorithms presented induce fault-tolerance and self-adaptivity. Performance evaluation based on simulations predicts a fast convergence time (1/33s for 64K nodes), exhibiting the promising properties of such self-stabilizing approach.

We pursue this work by implementing and evaluating the algorithms in the STCI runtime environment to validate the theoretical results.