BPF map lookup cost under cache pressure — raw benchmark data (Zen 4)

Raw benchmark output from a measurement study characterizing BPF map lookup cost under cache pressure on AMD Zen 4 (Ryzen 9 7900, 12C/24T, 2 CCDs). The sweep covers 5 map modes (`baseline`, `direct`-indexed array, `hash_map`, `lru_percpu`, `percpu_hash`) across 10 cardinalities (1K to 8M entries) and 3 access distributions (uniform, Zipf s=1.0, Zipf s=1.5), with 5 repetitions per configuration — 750 individual runs total. Each run captures PMU events stored under the shorter labels `cache_misses`, `ccd_l3_transfers`, and `l2_misses` in the `.profile` files. These correspond respectively to `ls_any_fills_from_sys.dram_io_near` (local DRAM fills — `cache_misses` is a historical perf-stat alias, not all-cache misses), `ls_any_fills_from_sys.remote_cache` (cross-CCD L3 fills via Infinity Fabric), and `l2_cache_req_stat.ic_dc_miss_in_l2` (L2 misses), attributed to the BPF program via `perf stat --bpf-prog`. One pre-sweep and five post-repetition pointer-chase DRAM-latency samples are included as thermal-stability evidence (1.35% drift across the full sweep). Hardware: AMD Ryzen 9 7900, DDR5-6000 CL50 (EXPO/Auto, tCL50-50-50-96), Mellanox MCX621102AC-ADAT ConnectX-6 Dx 2-port SFP28 25 GbE NIC (max 50 Gbps aggregate across both ports), ports linked back-to-back with a Ubiquiti UACC-DAC-SFP28-0.5M direct-attach copper cable. Kernel: Linux 6.18.20-lts. Isolation controls applied include CPU governor=performance with clock fixed at 5.49 GHz, SMT siblings offlined on CCD 0, C-states restricted to C1, NMI watchdog disabled, 6 RX queues pinned to CCD-0 cores 0–5, IRQ coalescing rx-frames=1, RPS/RFS disabled, NIC offloads off, traffic generator pinned to CCD 1 (cores 6–11), IPv6 disabled. Per-run profile window: 30 seconds after a 5–20 second adaptive warm-up. Tarball contents: per-run `perf stat` output (`.profile`), traffic-generator stderr (`.traffic.log`), BPF-loader stderr (`.loader.log`), per-run hit/miss stats (`.stats.json`), per-session hardware/kernel metadata (`metadata.json`, `dram_lat_measurements.json`), and the orchestration log (`run_bench.log`). Source and destination MAC addresses, network interface names, and the output mount path from the bench host were rewritten in `run_bench.log`, `.stats.json`, `.traffic.log`, and `.loader.log` using a parameterized sanitization script published with the analysis code. PMU counter values, timestamps, and run identifiers are byte-unchanged from what `perf stat --bpf-prog` emitted on the bench host. Code that produced this data and the analysis pipeline that consumes it: https://github.com/bpf-research/bpf-map-perf-pmu/