MIMIR
MIMIR is an online cache profiler that estimates hit-ratio curves by grouping recency information into buckets.
Bucketed recency estimates trade precision for lower profiling cost. Ghost identifiers or counting Bloom filters extend the estimated curve beyond resident capacity; a ghost match still requires fetching the payload. Filter false positives can distort predictions. [1]
The distributed construction assumes uniform random placement and equal per-server capacity. It is not an optimizer over arbitrary capacity splits. Compatible request ratios must pool hit and request counts, rather than average percentages. [1] [2]
Constructed cold-start LRU example: server A alternates two keys for 90 requests; B repeats one key for 10. One slot each gives 0 + 9 hits out of 100 (9%), not the unweighted 45%. Giving both slots to A yields 88/100 (88%) with routing unchanged. This is illustrative trace arithmetic, not a MIMIR reproduction. [3]
In the paper, curve accuracy is 1 minus mean absolute error across tested capacities. It is neither hit ratio nor a worst-capacity error bound. Two constructed 100-point curves can agree at 99 capacities yet differ by 100 percentage points at the last: 99% accuracy under that metric. [1] [3]
Relations
Limitations
- The basic model counts equal-size objects; byte-weighted extensions change the measurement unit. An estimated cache curve does not determine backend latency. [1]
Citations
- Saemundsson et al. Dynamic Performance Profiling of Cloud Caches. SoCC 2014, sections 2.3-2.5, 3.7 and 4. DOI: 10.1145/2670979.2671007.
- Prometheus. Recording rules: aggregate numerator and denominator separately before division.
- Hesela. Constructed LRU allocation and mean-versus-maximum curve-error controls.
Further reading
- Hesela. Exact cold-start LRU replay and curve-error arithmetic; no MIMIR algorithm implementation.
- Hesela. Exhaustive small-trace oracle, counter validation and curve-error regression tests.
- Hesela. Cache sizing, workload boundaries and hit-ratio accounting.
Sources & caveats
- The author-hosted 14-page SoCC 2014 paper was read beyond its abstract. Historical performance figures are not present-day guarantees.
- The examples use synthetic requests and constructed curves. They test interpretation, not MIMIR fidelity, concurrency or measured device performance.