mikrobenchmarks

Mikrobenchmarks are small-scale benchmarks that measure the performance of individual operations or components within a computer system. They aim to isolate a single aspect, such as a CPU instruction, a cache access, or an I/O primitive, in order to reveal low-level behavior of hardware and software.

They differ from macro benchmarks, which evaluate end-to-end performance on realistic workloads. Mikrobenchmarks are useful for comparing architectures, evaluating compiler optimizations, and guiding tuning, but their results are highly sensitive to microarchitectural features, cache state, power and frequency scaling, and operating-system scheduling.

Common targets include CPU instruction latency and throughput, memory-subsystem latency and bandwidth (L1/L2/L3 caches, main memory), and basic I/O operations such as disk accesses or network sends. Reported metrics typically include latency, throughput, bandwidth, and, for CPUs, instructions per cycle, as well as cache misses and stall cycles.

Popular tools and benchmarks in this category include lmbench for system and memory latencies, CoreMark and Dhrystone for CPU-oriented work, and STREAM for measuring memory bandwidth. The goal is reproducibility, often achieved with fixed input sizes and careful control of compilation and runtime conditions.

Limitations are substantial: results can be skewed by compiler optimizations, inlining, warm-up effects, CPU frequency scaling, and non-representative workloads. Mikrobenchmarks should be complemented by macro benchmarks and should report variance and confidence intervals.

In practice, Mikrobenchmarks help diagnose bottlenecks, compare hardware designs, and guide tuning, but they are most informative when interpreted alongside broader performance testing.