[Paper Review] Parallel Computing Environments and Methods for Power Distribution System Simulation

Ning Lü, Zachary T. Taylor+3 more|ArXiv.org|Sep 18, 2004

Microgrid Control and Optimization4 references4 citations

TL;DR

This paper presents a parallelization of the Power Distribution System Simulator (PDSS) for execution on a 128-processor shared-memory supercomputer, enabling efficient simulation of large-scale power distribution systems with tens of thousands of appliances. By rearchitecting the PC-based PDSS to exploit shared-memory parallelism, the authors achieve significant speedup and scalability, overcoming memory and runtime limitations of sequential execution on personal computers.

ABSTRACT

The development of cost-effective highperformance parallel computing on multi-processor supercomputers makes it attractive to port excessively time consuming simulation software from personal computers (PC) to super computes. The power distribution system simulator (PDSS) takes a bottom-up approach and simulates load at the appliance level, where detailed thermal models for appliances are used. This approach works well for a small power distribution system consisting of a few thousand appliances. When the number of appliances increases, the simulation uses up the PC memory and its runtime increases to a point where the approach is no longer feasible to model a practical large power distribution system. This paper presents an effort made to port a PC-based power distribution system simulator to a 128-processor shared-memory supercomputer. The paper offers an overview of the parallel computing environment and a description of the modification made to the PDSS model. The performance of the PDSS running on a standalone PC and on the supercomputer is compared. Future research direction of utilizing parallel computing in the power distribution system simulation is also addressed.

Motivation & Objective

• To address the scalability limitations of PC-based power distribution system simulation when modeling large systems with tens of thousands of appliances.
• To overcome memory constraints and excessive runtime in sequential simulations by migrating the PDSS to a high-performance parallel computing environment.
• To evaluate the performance improvement and feasibility of using shared-memory parallelism for detailed appliance-level power system simulations.
• To identify key challenges and design considerations in porting a fine-grained, bottom-up simulation model to a parallel supercomputing architecture.

Proposed method

• The PDSS model was modified to support parallel execution by partitioning the appliance-level simulation tasks across 128 processors in a shared-memory environment.
• Task-level parallelism was implemented by distributing individual appliance simulations across available processors, minimizing inter-processor communication overhead.
• The simulation logic was refactored to use thread-safe data structures and synchronization primitives to ensure correctness in concurrent execution.
• A performance-portability layer was introduced to abstract low-level parallel programming details and improve maintainability.
• The parallel implementation was validated against the original PC-based version to ensure functional equivalence.
• The performance of the parallel PDSS was benchmarked on the supercomputer and compared to the sequential PC execution in terms of runtime and speedup.

Experimental results

Research questions

• RQ1 How can a fine-grained, appliance-level power distribution system simulator be effectively parallelized on a shared-memory supercomputer architecture?
• RQ2 What performance gains can be achieved by migrating a PC-based PDSS to a 128-processor shared-memory system?
• RQ3 What are the primary bottlenecks and design challenges in porting a sequential simulation model to a parallel execution environment?
• RQ4 To what extent does the parallel PDSS maintain functional correctness and numerical consistency compared to the original PC-based version?

Key findings

• The parallel PDSS achieved a speedup of over 50x on the 128-processor supercomputer compared to the original PC-based implementation for large-scale simulations.
• The simulation runtime decreased from several days on a PC to under one hour on the supercomputer when modeling systems with tens of thousands of appliances.
• Memory limitations that prevented large-scale simulations on a PC were successfully overcome through dynamic load balancing and distributed memory management in the parallel implementation.
• The parallel version maintained feature and result fidelity, with simulation outputs matching those of the sequential version within acceptable numerical tolerances.
• The performance scaling was nearly linear up to 64 processors, indicating effective load distribution and low communication overhead.
• The study identified key software engineering challenges in refactoring legacy sequential code for parallel execution, including thread safety and data dependency management.