Bare-Metal Linux Profiling

Use a dedicated Linux host for actionable CPU and cache data. Docker Desktop results include a virtual machine and are useful only as development baselines.

Prerequisites

• perf, commonly provided by linux-tools or perf
• redis-cli and redis-benchmark
• taskset, commonly provided by util-linux
• Optional numactl for NUMA topology capture

The server and load generator must use disjoint CPU sets. Disable unrelated workloads and record CPU frequency policy, SMT state, kernel version, and NUMA topology.

Counter Profile

SERVER_CORES=2 CLIENT_CORES=4-7 MODE=stat ./scripts/profile_linux.sh

This records cycles, instructions, cache references/misses, branch misses, context switches, migrations, page faults, benchmark output, and INFO metrics.

Call Graph

SERVER_CORES=2 CLIENT_CORES=4-7 MODE=record ./scripts/profile_linux.sh

Inspect the generated report under profile/results/. Run at least five samples for each configuration and compare medians rather than selecting the best run.

Profile four shards with one pinned server CPU per shard and separate load-generator CPUs:

SHARDS=4 SERVER_CORES=2-5 CLIENT_CORES=8-11 MODE=both \
  REQUESTS=500000 CLIENTS=25 ./scripts/profile_linux.sh

The profile captures the whole process and writes one INFO snapshot per shard.

If perf reports a permission error, adjust kernel.perf_event_paranoid according to the host security policy. Do not grant elevated container privileges merely to collect benchmark numbers.