[Paper Review] Using Coloured Petri Nets for design of parallel raytracing environment

Štefan Korečko, Branislav Sobota|arXiv (Cornell University)|Mar 6, 2010

3D Modeling in Geospatial Applications8 references3 citations

TL;DR

This paper presents a Coloured Petri Net (CPN)-based model for designing and analyzing a parallel raytracing system in a cluster environment. It uses CPN tools for simulation-based performance analysis to evaluate load balancing and fault tolerance, revealing that more than ten nodes offer diminishing returns due to communication overhead and node failures in real-world conditions.

ABSTRACT

This paper deals with the parallel raytracing part of virtual-reality system PROLAND, developed at the home institution of authors. It describes an actual implementation of the raytracing part and introduces a Coloured Petri Nets model of the implementation. The model is used for an evaluation of the implementation by means of simulation-based performance analysis and also forms the basis for future improvements of its parallelization strategy.

Motivation & Objective

• To design a scalable and fault-tolerant parallel raytracing system using a cluster-based architecture.
• To model the raytracing process with Coloured Petri Nets (CPNs) for formal verification and performance simulation.
• To evaluate the performance of the implementation under varying cluster sizes and failure conditions.
• To identify bottlenecks in parallelization and guide future optimization of the load-balancing strategy.
• To establish a formal modeling framework that supports both analytical verification and simulation-based performance analysis.

Proposed method

• The system models raytracing as a state transition process using Coloured Petri Nets (CPNs), with tokens representing tiles, nodes, and job states.
• The model includes states for prepared, raytracing, failed, recovered, and computed tiles, with transitions for job assignment, raytracing, failure detection, and recovery.
• Raytracing time is computed from tile properties, while communication delay is modeled as an exponential random variable to simulate network and master-node contention.
• Failure detection is modeled via periodic checks, with a failure delay computed as a multiple of the check period, capped at one day.
• The model supports simulation of both ideal conditions (no failures) and real-world conditions (node failures, variable performance).
• Performance data, including raytracing duration and node usage, are logged upon completion of a scene for post-simulation analysis.

Experimental results

Research questions

• RQ1 How does the performance of the parallel raytracing system scale with increasing cluster size under ideal and real-world conditions?
• RQ2 What impact do node failures and recovery mechanisms have on overall raytracing duration and system reliability?
• RQ3 To what extent does communication overhead and load imbalance limit scalability beyond a certain number of nodes?
• RQ4 Can the CPN model effectively support both performance simulation and formal verification of the raytracing strategy?
• RQ5 What is the optimal number of nodes for efficient raytracing in the current implementation, given fault tolerance and communication costs?

Key findings

• In the ideal scenario, raytracing duration decreases with increasing cluster size, showing near-linear speedup up to a point.
• In the real scenario, raytracing duration is longer and less predictable due to node failures and communication delays, resulting in non-smooth performance curves.
• Using more than ten nodes in the current implementation yields diminishing returns, as communication overhead and failure handling reduce efficiency.
• The model successfully captures fault tolerance behavior, including failure detection, recovery, and node reintegration within a day.
• The CPN model enables effective simulation-based performance analysis and provides a foundation for iterative improvement of the parallelization strategy.
• The system's performance is highly sensitive to node failure rates and communication delays, which significantly impact total rendering time.