Performance versus Scalability

The primary goal of parallel programming is usually to improve performance, meaning reducing the time required to complete a task. However, performance alone is not enough to assess the quality of a parallel program. It’s equally important to consider scalability — the ability of the program to maintain efficiency as the number of computing resources or the problem size increases.

• Performance is about how fast a task is executed.
• Scalability is about how efficiently that performance improves with more resources.

To illustrate this distinction, imagine a team washing dishes at a restaurant. If only one person is washing, performance is measured by how many dishes they can clean per hour. Now, if the number of dirty dishes increases and more people are added to the team, scalability refers to how well the team maintains or improves efficiency. If the workload is evenly distributed and no one gets in the way, the team scales well. But if additional workers cause bottlenecks — waiting for space, tools, or coordination — efficiency drops, despite having more people. This mirrors what happens in poorly scalable parallel programs, where the overhead of managing multiple threads or processes outweighs the benefits of added resources.

• Performance analysis focuses on a detailed evaluation of a single configuration.
• Scalability analysis requires evaluating multiple configurations to understand how performance evolves.

Unfortunately, most performance profiling tools are not designed to assess scalability effectively:

• They require users to manually run and collect data across several configurations.
• They often generate overly detailed output, which can obscure key scalability insights.
• They typically lack intuitive visualizations that make scalability analysis easier.

Yet, scalability is a critical concern in modern high-performance computing. Large-scale machines are excellent for handling many tasks simultaneously, but that doesn’t mean they are efficient for solving a single large problem. Understanding and analyzing scalability is essential to ensure resources are used effectively.