Cloud Accelerator Scheduling for Physics Solver Partitioning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Legacy CAE software struggles to efficiently utilize GPUs due to re-architecting challenges, CPU-GPU communication overhead, and limited availability of cloud-based GPU resources, leading to suboptimal performance and high costs in cloud-native environments.
Innovation Solution
A cloud-based framework with a predictive scheduler and hardware-agnostic software layer dynamically partitions physics simulation software across multiple accelerators, optimizing resource utilization and managing heterogeneous networks to achieve supercomputing performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If legacy CAE software is re-architected to run entirely on GPUs, then computing performance is improved, but development cost and complexity increase substantially
Solution Approach 1:
The patent segments the CAE software into two distinct parts: a GPU-accelerated core for compute-intensive physics simulations and a CPU-based interface layer for data management, visualization, and user interaction. This segmentation allows the performance-critical portions to benefit from GPU parallelization while maintaining the existing CPU-based architecture for tasks that are better suited to general-purpose processing, thereby achieving performance improvement without requiring complete re-architecture of the entire software system.
Solution Approach 2:
The patent creates a universal software framework that can operate in multiple modes: fully CPU-based mode for compatibility with existing codebases, hybrid CPU-GPU mode for optimized performance, and fully GPU-based mode for maximum acceleration. This multi-functionality allows the software to adapt to different hardware configurations and user needs without requiring mandatory complete re-architecture, thus reducing development complexity while maintaining performance improvement opportunities.
2Productivity
If more GPU resources are allocated for simulation, then computational speed is improved, but resource cost and availability difficulty increase
Solution Approach 1:
The patent implements dynamic resource allocation where the system can automatically adjust the number and type of GPU resources allocated to simulations based on real-time performance monitoring and workload requirements. The framework dynamically partitions simulation data across multiple GPUs and adjusts computation distribution to optimize both performance and resource utilization, allowing users to achieve high computational speed without being locked into fixed, expensive hardware configurations.
Solution Approach 2:
The patent enables flexible parameter changes in resource allocation by allowing dynamic adjustment of precision levels (e.g., single vs. double precision floating-point), number of GPUs, and data partitioning strategies. Users can modify these parameters based on their specific computational needs, balancing between computational speed and resource cost, and adapting to different hardware availability scenarios without sacrificing simulation accuracy.
3Ease of operation
If data is transferred frequently between CPU and GPU, then data accessibility is improved, but communication overhead increases
Solution Approach 1:
The patent implements preliminary data preparation and pre-processing on the GPU before the main simulation computation begins. By transferring and preparing necessary data in advance during idle periods or initialization phases, the system minimizes the need for frequent data transfers during critical computation phases. This preliminary action reduces communication overhead while maintaining data accessibility when needed.
Solution Approach 2:
The patent introduces an intermediary memory hierarchy and data management layer that acts as a buffer between CPU and GPU memory systems. This intermediary structure includes shared memory regions and data caching mechanisms that reduce the frequency and impact of direct CPU-GPU data transfers. The intermediary layer manages data accessibility efficiently by providing fast access to frequently used data while minimizing unnecessary communication overhead.
4Ease of manufacture
If cloud-based GPU resources are used, then infrastructure cost is reduced, but resource acquisition difficulty and latency increase
Solution Approach 1:
The patent implements feedback mechanisms that continuously monitor cloud resource availability, performance, and cost parameters. Based on this feedback, the system dynamically selects optimal GPU resources from multiple cloud providers or regions, adjusting allocation decisions in real-time to minimize latency and maximize performance. The feedback loop also enables automatic scaling and load balancing across distributed cloud resources, reducing acquisition latency while maintaining cost-effectiveness.
Solution Approach 2:
The patent introduces a new dimension to resource selection by considering multiple cloud service providers, geographic regions, and hardware configurations simultaneously. Instead of relying on a single resource pool, the system performs multi-dimensional optimization across availability, performance, cost, and latency parameters. This dimensional expansion enables the system to find optimal resource allocations that balance infrastructure cost reduction with minimal acquisition latency, leveraging global cloud infrastructure diversity.
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.