Ricerc@Sapienza

Nowadays, many-core platforms are the de-facto reference architectures for current and future processors. This is mainly due to physical limitations that do not allow to improve the sequential performance of an individual CPU-core.
Contextually, such a scaled up hardware parallelism has given rise to innovative coordination mechanisms that overstep the classical ones based on lock-protected critical sections.
In particular, non-blocking coordination algorithms make threads advance their computation regardless of whether other threads are blocked or faulty, thus providing definitely improved scalability in the management of shared data under high-contention.

However, literature results have demonstrated that, given a coordination problem, non-blocking algorithms solving it do not provide a one-size-fit-all solution, e.g., given that they may induce retries of specific algorithmic steps or because of algorithm complexity aspects.
In this context, our research objective will be the to design and implement non-blocking coordination algorithms which can be able to adapt their run-time behavior according to the workload profile, the variable intensity of shared-data access contention, and the peculiarities of the underlying hardware platform (such as the organization of the memory hierarchy).
The experimental assessment of our achievements will be carried out along two main directions: (i) by using several kinds of concurrent applications as test-beds (HPC instances, OSes, hypervisors, soft real-time systems); (ii) by testing our solutions on different platforms (multi-core, many-core, GPU, cloud-based).