Monitoring communication channels on a shared memory multi-processor system on chip

Abstract

To meet performance requirements, streaming applications have been mapped to Multi-Processor System on Chip (MPSoC). The Kahn Process Network (KPN) paradigm is sufficient when dealing with pipeline parallelism, but such pointto-point channels are impractical in the presence of massive task farm parallelism.Multi Writer Multi Reader (MWMR) channels generalize KPN in such a way that multiple writers and multiple readers access the same channel. They are implemented as software channels stored in on-chip memory to accommodate access by hardware and software tasks alike. The price to pay for this implementation is increased traffic to and from memory.Typical representatives are telecommunication applications which may treat hundreds or thousands of flows at a time, where the same chain of treatments is applied to every packet. The latency for this treatment depends on the packet's content, and can thus not be foreseen. Among multiple tasks which access a MWMR channel, the time to obtain a lock is variable. In consequence, fill states of MWMR channels vary heavily and it is crucial to monitor it in order to detect potential bottlenecks. We show how this can be done early in the design process by using SoCLib/DSX.

System-Level Design and Virtual Prototyping of a Telecommunication Application on a NUMA Platform

“…They are based on the VCI_logger module proposed in SoCLib. Figure 4 shows how this spy mechanism, described in [12], is now integrated in the deployment diagram. These spies are represented in the form of magnifying glasses.…”