The global objective of this challenge consists in providing i) a simulation infrastructure of the global manycore platform and ii) all the device models necessary for simulating the platform. This challenge will focus on accurate low-level simulations of new photonics devices to extract accurate models and to integrate them into the manycore simulation infrastructure, with the objective to produce high-level models for fast design space exploration in Challenge 1.
Participants
INRIA + INL
Description
In the context of 3D architecture simulation, the overall manycore architecture (i.e. including ONoC) will be modeled at the system level in order to allow system designers to simulate the execution of realistic data-intensive applications implemented on the platform. Based on the simulation results, information on overall execution performance (e.g. latency and throughput) and power consumption will allow exploration of the design space (Challenge 1). For this purpose, two sets of models will be considered:
- Low abstraction-level modeling for accurate simulations of optical communications (thousands data bits) relying on new optical devices and new optical functions, respectively, defined in Challenges 2 and 4. For this part, a hybrid simulation tool providing very accurate results about performance and energy consumption will be necessary. From these simulation results, accurate models of optical components will be defined to be used as parameters of the high-level simulation defined below. Accurate models for electrical to optical interfaces will also be extracted to complete the platform simulation infrastructure.
- High abstraction-level modeling for system-level simulations of the overall platform, including the optical interconnect and the electrical layers (several Gbits of data communication will thus be considered to realistically represent the communication activity), which is thus suitable for design space exploration in Challenge 1. To address this point, a heterogeneous simulation tool must be defined to support both high-level electrical and optical simulations. The important point here is to start from an existing simulation infrastructure of a 2D manycore platform and to extend it to 3D and optical. Third dimension and ONoC models will be included into the simulator and characteristics of elementary optical components will be included as parameters of the NoC. The choice of an existing manycore simulation infrastructure will be an important milestone of this challenge.
The aim of the low-level models is to characterize power and performances of simple optical communication (e.g. a direct optical link) relying on the new photonic devices and new optical function developed at FOTON. For this purpose, heterogeneous and hybrid (optical, analog, and digital) simulation models will be defined and the Verilog-A models already available at INL will be updated and extended accordingly. The inputs of the models will be the device characteristics (e.g. laser threshold current) and the output is a report containing performance information (e.g. power and latency) according to device integration characteristics (e.g. link size, number of crossed micro-resonators).
The high-level platform model will be based on several layers of the 3D platform including processor and NoC layer, reconfigurable and hardware accelerator layer, memory layer; photonics communication layer; and the 3D network-on-chip. Specific attention will be given to the model of the photonics layer, which will mostly rely on abstracting the low-level analog and optical models with power and delay annotations in the actual 3D-NoC model to allow for scalability of the simulation performance. The result will be, e.g., a single SystemC annotated model. With these different layers, the project can cover all possible communication profiles between all hardware and software components. Typical applications defined as parallel algorithms and with several communication protocols will be used to demonstrate the potential of optical interconnects with respect to classical electrical networks on chip. Therefore, a simulator will be necessary to validate and evaluate the complete manycore platform and will thus be developed in this project.
Many simulators already exist in the literature and have specific focus on NoC or MPSoC. Because the project aims at simulating hundreds of cores interconnected through an ONoC, we will focus on the use of parallel simulators that allow drastic reduction of simulation time mandatory for design space exploration (Challenge 1). Most advanced parallel simulators are Graphite[1] from MIT, Sniper[2] from Ghent, Rabbits[3] from TIMA, and gem5[4]. An internal project in INL already starts to evaluate the simulator the most suitable for the accuracy / execution time tradeoffs. Exploration of the different simulation tools will be further conducted and the simulation environment will be selected at the beginning of the project. We will then propose some extensions to support 3D and optical communications through the NoC to help the designer to explore performance and energy consumption of the platform.
The main inputs of this platform modelling task are i) the characteristics of the novel photonic devices proposed by FOTON and ii) the analytical models of the corresponding new functions. Indeed, these characteristics can ease or allow the execution of new operations at system level (e.g. broadcast) and the analytical models will allow evaluating their impacts on performance and power consumption. Since design parameters require modifications of the interconnects, a strong interaction with the ONoC architecture exploration part will be achieved.
