Benchmarking Redis In-Memory Eviction Strategies (7439)

Technical Overview

Engineering breakdown of Benchmarking Redis In-Memory Eviction Strategies (7439). Bare-metal hardware performance requires isolated kernel parameters, NUMA‑aware memory allocation, and precise control over Redis' maxmemory-policy settings. This article walks through a reproducible methodology for measuring latency, throughput, and eviction accuracy across the seven built‑in policies: noeviction, allkeys-lru, volatile-lru, allkeys-lfu, volatile-lfu, allkeys-random, and volatile-random.

Testbed Configuration

• CPU: 2× AMD EPYC 7763 (64 cores each), SMT disabled
• RAM: 1 TiB DDR4‑3200, 8‑channel, interleaved
• OS: Ubuntu 22.04 LTS, kernel 6.5, transparent_hugepage=never
• Redis: 7.2.4 compiled with JEMALLOC and THP disabled
• Network: 25 GbE, RSS enabled, IRQ affinity pinned

Workload Generator

A custom Go client (v1.22) drives a mixed read/write pattern (70 % GET, 30 % SET) with a Zipfian key distribution (θ=0.99) to emulate realistic cache pressure. Each run lasts 300 seconds, preceded by a 30‑second warm‑up. Metrics are collected via redis-cli --stat, perf stat, and eBPF probes on redisServer.evict.