Cloud Physics Solver Partitioning Across Heterogeneous Accelerators
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Legacy CAE software struggles with limited apportionment of physical simulation tasks on GPUs, leading to substantial communication overhead, inefficient memory utilization, and difficulty transitioning to cloud-native environments, with limited access to cost-effective GPU resources.
Innovation Solution
A cloud-based framework that dynamically partitions physics solver code across multiple accelerators using a predictive scheduler and hardware-agnostic software layer, optimizing resource utilization and parallel execution, and asynchronously managing compute and storage resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If legacy CAE software is executed on GPUs with limited apportionment, then some parallel computation capability is achieved, but substantial communication overhead occurs between CPU and GPU
Solution Approach 1:
The patent segments the physics simulation software into distinct CPU-executed portions (preprocessing, postprocessing, control logic) and GPU-executed portions (linear solver kernels). This segmentation enables optimal placement of computational tasks on the appropriate hardware platform, minimizing unnecessary data transfers and communication overhead while maximizing parallel computation capability.
Solution Approach 2:
The patent introduces a hardware-agnostic software layer as an intermediary between the legacy CAE software and the GPU accelerator. This intermediary layer handles the interface between CPU and GPU, managing data transfers and coordinate transformations efficiently, thereby reducing communication overhead while enabling full utilization of GPU parallel processing power.
2Productivity
If legacy CAE software is re-architected entirely for GPU execution, then optimal GPU performance is achieved, but substantial effort and cost are required
Solution Approach 1:
Instead of complete re-architecture, the patent segments the existing legacy CAE software to identify and extract only the computationally intensive linear solver portions that benefit most from GPU acceleration. This selective approach achieves significant performance improvement while avoiding the substantial effort required to re-architect the entire software system.
Solution Approach 2:
The hardware-agnostic software layer serves as a bridge that allows legacy CAE software to execute on GPUs without complete re-architecting. This intermediary layer handles the complexity of GPU-specific optimizations and data formats, enabling legacy software to leverage GPU acceleration with minimal modification to the original codebase.
3Productivity
If mixed-precision arithmetic is used in GPUs, then bandwidth requirements decrease and performance improves, but software compatibility issues arise
Solution Approach 1:
The patent implements parameter changes in the data representation layer, enabling the software to dynamically select between single-precision and double-precision floating-point formats based on the specific computational requirements. This allows the system to exploit mixed-precision arithmetic for improved performance while maintaining compatibility with existing CAE software through the hardware-agnostic interface.
4Adaptability or versatility
If cloud-based GPU resources are leased on-demand, then resource flexibility is improved, but access to desirable computational resources becomes difficult
Solution Approach 1:
The patent implements a predictive scheduler that performs preliminary actions by anticipating future computational needs and pre-reserving GPU resources before they are actually required. This proactive approach ensures that desirable computational resources are secured in advance, avoiding the difficulty of accessing available resources on-demand while maintaining the flexibility of cloud-based deployment.
Data Source
AI summary
A cloud-based framework dynamically utilizes a distributed pool of accelerators to parallelize calculations of physical simulation (physics) solver code partitioned across multiple accelerators and compute nodes of one or more virtual data centers in a virtualized computing environment. Multi-level partitioning logic of the framework partitions an input data set of the physics solver code into code groups configured to run on the accelerators using a “hardware agnostic” software layer that abstracts differences in processing architectures to allow targeting of different types of accelerators. A predictive scheduler interacts with the multi-level partitioning logic to locate and predictively reserve the accelerators within the pool, dynamically access and utilize the accelerators when needed, and then promptly release them upon completion of the calculations. The framework is configured to efficiently use bandwidth/compute capacity of the accelerators for physics solver code calculations asynchronously and in cooperation with general-purpose processing units as needed and on user demand.