Section: New Results

NewMadeleine generic multi-threading

The PIOMan progression engine utilized in NewMadeleine used to rely on the Marcel specific multi-threading library, with dedicated hooks and close co-operation between libraries. It restricted the target platforms and applications, and was considered as a constraint by users. We have designed mechanisms to make communication progress without hooks in the thread scheduler, able to run on any system with a pthread library. We have re-written PIOMan from the ground up to implement these mechanisms, and based on lock-free structures, with scalability in mind. A proof-of-concept port to the Intel Xeon Phi has been implemented in cooperation with the University of Tokyo, using the DCFA (Direct Communication Facility for manycore-based Accelerators) library to access InfiniBand boards from the Xeon Phi.