Engineering Simulations on Public Grids via Software Message Passing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Public grids, such as AWS and GOOGLE CLOUD, lack high-performance networks necessary for advanced engineering simulations, which are typically required for high-performance computing applications, limiting their use for simulations that demand extensive message passing between machines.
Innovation Solution
Implementing strategies to efficiently perform engineering simulations on public grids with limited network capability by minimizing message passing, optimizing thread efficiency, and maximizing cache performance, including geometric partitioning, local assembly, and novel matrix solvers, to achieve scalability and computational efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If high-performance network hardware is used to enable advanced engineering simulations, then computational capability and message passing efficiency are improved, but system cost increases significantly
Solution Approach 1:
The patent uses software-based message passing implementations that replicate the functionality of high-performance networks through standard network infrastructure. Multiple software layers (MPI implementations, communication libraries) copy and simulate the behavior of dedicated high-performance network hardware, enabling simulations on public grids without requiring expensive specialized network equipment.
Solution Approach 2:
The patent changes the operational parameters of standard networks to optimize them for simulation workloads. This includes adjusting message passing protocols, implementing aggregation strategies, and tuning communication patterns to reduce the effective bandwidth requirements, allowing standard networks to achieve performance levels previously only available from high-performance hardware.
2Device complexity
If standard networks are used for engineering simulations, then system cost is reduced, but message passing efficiency and scalability deteriorate
Solution Approach 1:
The patent segments the simulation workload and communication patterns to reduce the burden on any single network connection. By dividing the computational domain into smaller sub-domains and optimizing the message passing between them, the system achieves better scalability on standard networks, as each individual connection handles less traffic while the overall system maintains parallel efficiency.
Solution Approach 2:
The patent introduces software intermediaries (communication libraries, message passing interfaces, aggregation layers) that mediate between the simulation algorithms and the standard network infrastructure. These intermediaries optimize communication patterns, buffer and aggregate messages, and translate high-level simulation requirements into network-efficient operations, thereby improving scalability without requiring high-performance hardware.
3Speed
If high-performance networks are deployed to support cluster computing, then message passing speed is improved, but the overall system cost exceeds that of the computers themselves
Solution Approach 1:
The patent employs software-based communication solutions that are inexpensive and can be rapidly deployed and modified compared to physical network hardware. These software layers act as disposable, easily replaceable components that provide high-performance message passing capabilities through standard networks, avoiding the need for expensive, difficult-to-modify high-performance network infrastructure.
Data Source
AI summary
A method and system provide performance computing on a public grid. A computational mesh to solve a received engineering simulation (that has computations/processes) is computed. The mesh divides the CAD model up spatially into elements where conservation equations are formed. Mesh vertices identify nodes where field variables are defined. Each process owns a node/element and is assigned to a mesh partition. A solver application installed on each computer and each computer CPU spawns a thread to each CPU core. A matrix is assembled by each thread that computes numerical operators (that are stored as a coefficient matrix) that represent the conservation equations. The matrix is solved by solving the equations represented by the coefficient matrix. The solving yields field vectors that update the field variables at the nodes. The engineering solution is rendered based on the updated field variables.