Abstract
The EURETILE project (2010-2012) addresses the scalability and reliability challenges of exascale computing by proposing a hierarchical hardware and software framework inspired by the brain's cortical structure. This paper summarizes the first three years of the project, highlighting the development of a unified tool-chain, custom network processors, and a systemic fault-tolerance paradigm. Key results include the QUonG prototype achieving 32 TFlops and significant speedups in neural simulation and Lattice QCD benchmarks.
1. Introduction
As computing systems scale toward the exascale, they encounter significant "communication" and "reliability" walls. Traditional many-core architectures often fail to efficiently manage thousands of concurrent processes or handle the inevitable hardware failures inherent at such scales. The EURETILE project (European Reference Tiled Experiment) was established to bridge this gap by creating a holistic programming and hardware paradigm capable of managing dynamic workloads and providing systemic fault tolerance.
2. The EURETILE Hierarchical Model
Drawing inspiration from the organization of the human brain, EURETILE utilizes a three-level hierarchy to organize many-tile systems:
- Level 1: Elementary Tiles (Cortical Columns): These serve as the basic building blocks, consisting of multi-processors with RISC cores, VLIW DSPs, and custom accelerators.
- Level 2: Tile Arrays (Cortical Areas): Multiple hardware tiles are interconnected via a custom Distributed Network Processor (DNP), facilitating efficient local communication.
- Level 3: The Full System (Neo-cortex): The entire architecture is managed by the Distributed Application Layer (DAL), which enables multiple dynamic applications to share hardware resources seamlessly.
To support this hardware, the project developed a Y-Chart tool-chain. This software flow automatically maps high-level Kahn Process Networks (KPN) onto the hardware, generating the necessary Hardware-dependent Software (HdS) to bridge the gap between application logic and physical execution.
3. Systemic Awareness and Fault Tolerance
A core innovation of EURETILE is the Local Fault Monitor (LO|FA|MO). Rather than relying on traditional "checkpoint and restart" methods, EURETILE treats hardware faults as standard system events. Local monitors detect failures in real-time and propagate this information through the hierarchy. This "systemic awareness" allows the middleware to automatically remap tasks from failed tiles to healthy ones, ensuring continuous operation without application crashes.
4. Benchmarking and Performance
The architecture was validated using three "Grand Challenge" benchmarks:
- DPSNN-STDP: A distributed cortical simulation used to test extreme parallelism and bio-inspired communication.
- Lattice Quantum Chromo Dynamics (LQCD): A high-performance computing (HPC) benchmark requiring massive numerical throughput and low-latency interconnects.
- Embedded DSP Benchmarks: Including MJPEG decoders and FFTs to test dynamic application mapping.
4.1 Key Results
- Hardware Prototyping: The QUonG prototype, a 16-node system with 16,000 cores, achieved a peak performance of 32 TFlops.
- ASIP Acceleration: A custom Application-Specific Instruction-set Processor (ASIP) for LQCD kernels performed 33x faster than a standard 32-bit CPU and 4x faster than a high-end VLIW DSP.
- Interconnect Efficiency: The APENet+ board and DNP provided 34 Gbps of bandwidth per direction with significantly reduced access latency compared to commercial alternatives.
- Simulation Speed: The parSC parallel SystemC kernel delivered a 7.77x speedup, significantly accelerating the design cycle for complex many-tile systems.
5. Conclusion and Future Work
The first three years of EURETILE successfully demonstrated the viability of a brain-inspired, hierarchical approach to many-tile computing. While the individual components—DAL, HdS, and DNP—are functional, future work will focus on deeper system-level integration. Key challenges remain in seamless task migration across non-uniform memory access (NUMA) architectures and the refinement of virtualization layers through the project's in-house DNA-OS. Subsequent phases will focus on extensive quantitative validation to identify and eliminate remaining middleware and hardware bottlenecks.
Summary updated