CONCEPT Cited by 1 source

Strong scaling

Definition

Strong scaling measures how much faster a fixed workload completes as you add more compute resources. It is the classic parallel-speedup question: fix the work, add machines, observe the ratio. In contrast, weak scaling holds per-node work constant and measures how total throughput grows.

The theoretical limit is set by Amdahl's law: if fraction s of the work is serial, speedup on N nodes caps at 1/(s + (1−s)/N). For scan-dominated analytical queries where most work is embarrassingly parallel reads + local aggregation, practical speedups approach N (linear).

Trade-offs

• Near-linear strong scaling requires even data distribution; skew concentrates work on hot nodes and caps speedup below ideal.
• Coordination overhead (shuffle, merge) grows with node count, so returns diminish for queries that are join- or window-heavy.
• Strong scaling does not tell you about throughput scaling (how many concurrent queries a cluster handles) — the two must be measured separately.

Seen in

• sources/2026-07-29-redpanda-single-query-scaling-in-redpanda-sql: Redpanda SQL demonstrates 7× speedup on 8 nodes for a 1 TB high-cardinality GROUP BY — close to the ideal diagonal for scan-dominated work.
• TPC-H Power Test (single-stream) is the industry-standard strong-scaling benchmark for analytical engines.