Auditorium (Building IMAG)
16 November 2018 - 10h30
Parallel Code Generation of Synchronous Programs for a Many-core Architecture (Phd Defense)
by Amaury Graillat from VERIMAG/KALRAY
Abstract: Most critical systems are subject to hard real-time requirements.
These systems are more and more complex and the computational
power of the predictable single-core processors is no longer
sufficient. Multi- or many-core architectures are good
alternatives but interferences on shared resources must be taken
into account to avoid unpredictable timing effects. For
many-core, the Network-on-Chip (NoC) must be configured such that
deadlocks are avoided and a tight Worst Case Traversal
Time (WCTT) of the communications can be computed. The Kalray
MPPA2 is a many-core architecture with good timing properties.
Dataflow Synchronous languages such as Lustre or Scade are widely
used for avionics critical software. In these languages, programs
are described by networks of computational nodes. We introduce a
method to extract parallel tasks from synchronous programs. Then,
we generate parallel code to deploy tasks on the chip and
implement NoC and shared-memory communications. The generated
code enables traceability. It is based on a time-triggered
execution model which relies on a static schedule and minimizes
the memory interferences thanks to usage of memory banks. The
code enables the computation of a worst case execution time bound
accounting for the memory interferences and the WCTT of NoC
transmissions. We generate a configuration of the platform to
enable a fair bandwidth attribution on the NoC, bounded
transmissions through the NoC and clock synchronization. Finally,
we apply this toolchain on avionic case studies and synthetic
benchmarks running on 64 cores.
