System-level design for communication-centric task farm applications

Abstract

Abstract-Massively parallel applications such as telecommunication and video streaming have the particularity that a large proportion of the time is spent on accessing communication channels between the tasks, due to contention on the on-chip interconnect. Moreover, the analysis of a given task deployment is often fastidious. Thus, we propose to extend an existing easy-to-use System-level Design methodology to task farm applications. The contribution first concerns adding relevant SysML modeling elements to take into account application code, hardware platforms and deployment constraints. Secondly, new modeling elements -including access techniques to communication channels -must be given a semantics in order to transform models into a well-defined SystemC virtual prototyping MPSoC platform. A telecommunication application serves as an example.

High-level modeling of communication-centric applications: Extensions to a system-level design and virtual prototyping tool

“…In [1], we have shown how to rely onnon-deterministic operators of timed automata in order to capture multi-writer multi-reader (MWMR) communications as already explained in in [26] i.e. any number of reader or writer tasks can access simultaneously to these channels.…”

“…Many-to-many communications are not common practice. Block hierarchy can be used to express the fact that several blocks write to the same channel, as shown in previous work [1]. Figure 3 shows a typical task-farm application modeled as AVATAR block diagram.…”

“…Our previous work [1,2] enabled the multi-level modeling of high throughput task-farm applications, by providing a formal basis for non-deterministic channel accesses and extended TTool to model NUMA architectures and generate NUMA virtual prototypes. However, while streaming applications strongly rely on I/O elements, it was not possible to model these elements.…”

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

“…The classification tasks are represented in an hierarchical manner but, as the behavior is described much more precisely (packet inspection, non-arbitrary choice of the priority queue), they are far more complex than in [11]. All classification tasks write into one of three priority queues, corresponding to three information signals exchanged on the central channel in the SysML block diagram.…”

“…In [11], we proposed a SysML based design methodology for task-farm applications on a flat interconnect. Only a simplified version of the application was handled, abstracting from I/O and internals (such as generation of descriptors, feedback of addresses, inspection of packets).…”