Scalability
How well a system's performance improves as resources or problem size grow.
Definition
Scalability is the ability of a system to handle increased work by adding resources. In HPC, strong scaling holds the problem size fixed and adds processors; weak scaling grows the problem in proportion to the processors.
Reporting parallel efficiency, the fraction of ideal speedup actually achieved, is more informative than speedup alone, because it exposes where added processors stop paying off. Honest scaling studies vary processor count over a wide range rather than citing a single favorable point.
Strong and weak scaling answer different questions, and reporting the wrong one flatters a result: strong scaling asks how much faster a fixed problem runs on more processors, while weak scaling asks whether a proportionally larger problem holds its runtime. Communication overhead, serial bottlenecks, and load imbalance all erode scaling, so honest studies vary processor counts widely and report parallel efficiency rather than a single favorable data point.
Why scaling breaks down
- Communication overhead grows with processor count.
- Serial bottlenecks dominate, per Amdahl's law.
- Load imbalance leaves some processors idle.
- Memory or bandwidth limits saturate.
Why it matters
Scalability determines whether throwing more hardware at a problem actually helps. A code that scales well can exploit the largest machines; one that does not wastes them. Measuring and improving scaling is a core HPC discipline.
Fusion connection
Good parallel scalability is what lets Kronos raise simulation resolution or run more design cases by using more of a cluster, rather than being stuck at a fixed capability.