Belenos: Bottleneck Evaluation to Link Biomechanics to Novel Computing Optimizations

TL;DR

Belenos addresses the bottlenecks in biomechanical finite element simulations by providing a comprehensive architectural profiling of FEBio workloads and exploring microarchitectural sensitivities with gem5. The approach combines real-system VTune analysis to identify bottlenecks (notably backend-bound memory behavior) with full-system simulations to quantify how core frequency, cache capacity, pipeline width, memory queues, and branch predictors impact performance. Key findings show that irregular memory access and sparse computations dominate performance, with large models like the ocular eye case placing sustained pressure on memory hierarchies; improvements from simply increasing frequency or pipeline width are limited. The study argues for architecture-aware co-design, including near-memory processing and domain-specific accelerators, to accelerate iterative biomechanics workflows and enable higher-fidelity simulations with clinically relevant turnaround times.

Abstract

Finite element simulations are essential in biomechanics, enabling detailed modeling of tissues and organs. However, architectural inefficiencies in current hardware and software stacks limit performance and scalability, especially for iterative tasks like material parameter identification. As a result, workflows often sacrifice fidelity for tractability. Reconfigurable hardware, such as FPGAs, offers a promising path to domain-specific acceleration without the cost of ASICs, but its potential in biomechanics remains underexplored. This paper presents Belenos, a comprehensive workload characterization of finite element biomechanics using FEBio, a widely adopted simulator, gem5 sensitivity studies, and VTune analysis. VTune results reveal that smaller workloads experience moderate front-end stalls, typically around 13.1%, whereas larger workloads are dominated by significant back-end bottlenecks, with backend-bound cycles ranging from 59.9% to over 82.2%. Complementary gem5 sensitivity studies identify optimal hardware configurations for Domain-Specific Accelerators (DSA), showing that suboptimal pipeline, memory, or branch predictor settings can degrade performance by up to 37.1%. These findings underscore the need for architecture-aware co-design to efficiently support biomechanical simulation workloads.

Figures (12)

• Figure 1: Belenos workflow including the FEBio lifecycle shown in three stages, and the VTune to profile the performance characteristics of Stage 2, which consists of solving the biomechanical models.
• Figure 2: Top-down pipeline breakdown.
• Figure 3: Breakdown of front-end (FE) and back-end (BE) pipeline stalls across FEBio workloads. Bad speculation is excluded from the plots as its contribution is negligible.
• Figure 4: Prevalence of common function types within the top 5% of clockticks across workloads. Dot color reflects the fraction of top hotspots from each function category.
• Figure 5: Simulation time and model size for the eye model and test suite models. The models are organized by size and mapped onto a logarithmic scale.