Multicast On-chip Traffic Analysis Targeting Manycore NoC Design

Abstract

The scalability of Network-on-Chip (NoC) designs has become a rising concern as we enter the manycore era. Multicast support represents a particular yet relevant case within this context and has been the focus of different research efforts, mainly due to the poor performance of NoCs in the presence of this increasingly important type of traffic. However, most of the proposed schemes have been evaluated using synthetic traffic or within a full system, which is either unrealistic or costly. While traffic models would allow to better assess their performance, existing proposals do not distinguish between unicast and multicast flows and often are bound to a given number of cores. In this paper, a trace-based multicast traffic characterization is presented with the aim to provide guidelines for the modeling of multicast communications in manycore settings. To this end, the scaling trends of aspects such as the multicast traffic intensity or the spatiotemporal injection distribution are analyzed. The novelty of this work resides both on its scalability-oriented approach and on the use of correlation metrics to evaluate potential prediction opportunities.

Characterization and modeling of multicast communication in cache-coherent manycore processors

“…However, to the best of the authors' knowledge, no tools are available for the analysis and modeling of multicast traffic. Different works have characterized or modeled inter-processor communication in moderately-sized shared memory processors (see [2] and references therein), as well as in message passing clusters or supercomputers [3][4][5]. However, none of them has analyzed how collective communication scales for a wide set of representative applications and architectures at the CMP scale.…”

“…The paper is a direct extension of previous work by the authors [2]. The original contribution is augmented as follows:…”

“…Our previous work in [2] contains a comprehensive view of such characterization efforts, including the seminal papers that explore the SPLASH-2 and PARSEC benchmarks [6][7][8]. Their main downturn, though, is that those works do not distinguish between unicast and multicast in most cases.…”

“…Therefore, the size of these messages can be easily inferred from it and, as further explained in [2], taken into account in the NoC design process.…”

“…For instance, our previous work assumes a mesh NoC and yields slightly lower Hurst exponents in almost all cases. We refer the reader to [2] for more details.…”