go-cache-benchmark

Benchmark accompanying the paper «Адаптивное локальное кэширование в микросервисных системах с учетом особенностей сборки мусора» (Adaptive TinyLFU + GC-Aware Cost).

What is measured

Compared systems (all under the same byte budget):

Name in results	System
GCAware	proposed method, λ = 0.7 (internal/gcaware)
Adaptive_TinyLFU	the same implementation with λ = 1 (no cost term in V(x))
Ristretto	dgraph-io/ristretto v0.2.0, cost = object size
BigCache	allegro/bigcache v3, hard limit = budget

Go-Cache is not part of the comparison: it has no size limit and cannot be compared under an equal memory budget.

Series (one row per configuration and cache in results.csv):

Series	Configurations	Workload
size	budget 64 MB … 1 GB	5M keys, uniform, 10M requests, 1 KB objects
value	object 256 B … 64 KB	500k keys, Zipf s = 0.95, 2M requests, 512 MB
skew	s = 0.80 … 0.99	1M keys, 5M requests, 1 KB objects, 256 MB
mixed	—	200k keys, Zipf s = 0.9, 3M requests, 16 MB; sizes 256 B / 1 KB / 16 KB / 64 KB with weights 70/20/7/3
lambda	λ = 0 … 1 (GCAware only)	same as mixed

Protocol (see internal/bench/runner.go): look-aside access — Get, and on a miss a fresh object of the key's size is allocated (simulated fetch from the source) and Set. Latency covers the cache calls only. GC/allocation metrics are runtime.MemStats deltas over the request loop; gc_pause_us is the mean pause per GC cycle (ΔPauseTotalNs / ΔNumGC); alloc_bytes_per_op is ΔTotalAlloc per request (independent of speed). byte_hit_ratio is the share of requested bytes served from the cache (hit_ratio counts requests). memory_mb is the HeapInuse delta after a forced GC with the cache alive, relative to a baseline taken before the cache is constructed (live_mb is the same for HeapAlloc). Zipf ranks are sampled exactly (inverse CDF) and mapped to keys by a random permutation. Each run uses its own seed (the run number).

Running

Requires Docker with Compose. A full run takes several minutes per seed and needs about 6 GB of free RAM (the 1 GB budget configuration).

./run-benchmarks.sh                       # 30 runs -> results/run_XX, results/summary.csv
NUM_RUNS=3 ./run-benchmarks.sh            # fewer runs
BENCH_SCALE=0.05 NUM_RUNS=1 ./run-benchmarks.sh   # 1-minute smoke test

results/host.txt records the machine (CPU, RAM, OS, Docker); each results/run_XX/env.json records the Go version, GOMAXPROCS and GOGC.

Unit tests: cd bench-runner && go test ./...