MPKI

Misses Per Kilo Instructions (MPKI) is a performance metric used in computer architecture to quantify cache and translation lookaside buffer misses relative to the number of instructions executed. It is defined as the number of cache or TLB misses divided by the number of instructions executed, scaled to thousands of instructions, i.e., MPKI = misses / (instructions / 1000). MPKI is commonly reported for various memory hierarchy components, such as L1 data cache MPKI, L2 cache MPKI, instruction cache MPKI, and TLB MPKI (for both ITLB and DTLB).

MPKI is used to characterize memory locality and to compare how different workloads or architectural configurations interact with the memory system. A lower MPKI indicates fewer misses per thousand instructions and often correlates with better memory locality and potentially higher performance, though it is not a direct measure of total execution time.

Measurement of MPKI typically relies on hardware performance counters available in modern processors. Tools that access these counters (for example, performance profiling and tracing suites) count cache and TLB misses and the total number of instructions executed for a specified region of code or workload. Care must be taken to ensure consistent measurement windows, representative workloads, and to account for overhead or variation between runs.

Limitations of MPKI include its abstraction over complex memory behavior, sensitivity to hardware specifics (cache sizes, associativity, prefetchers), and the fact that it does not directly capture stall cycles or latency. It is most useful as a comparative metric for similar architectures and workloads, rather than as an absolute predictor of overall performance.