loaded-latency: add mode-2 multipass random chain (v1)

[realmzstevenmiao]

Merge the 8-pass random Hamiltonian chain randomization (mode 2) Defeating Neoverse V2 L2 prefetchers that learn the single-chain pattern used by mode 1.

Changes:

• args.c: add -R/--lat-randomize-mode to select randomization mode (1 = pair-swap shuffle, 2 = 8-pass random chain). -r remains shorthand for mode 1.

• memlatency.c:

  • Add make_multipass_chain() building LAT_CHAIN_PASSES (=8) independent Hamiltonian cycles via ptr_t slots per cacheline, concatenated head-to-tail into one closed loop. Each pass uses an independent Fisher-Yates random visitation order and a different slot offset within each cacheline, defeating stride, next-line, and short-history temporal prefetchers.
  • Refactor make_pairswap_chain() to use an external order[] array (mirroring mode 2). Drop per-node order/index bookkeeping fields; nodes now carry only ->next.
  • Shrink local node_t in lat_initialize() accordingly; union is now { void *next; ptr_t ptrs[LAT_CHAIN_PASSES]; } + cacheline pad.
  • Replace dead #if 0 debug block with #ifdef LAT_DEBUG_CHAIN that derives cacheline index/slot from the pointer and buffer base.

  ./smoke-mode1-vs-mode2.sh
  Using CPU 31 on NUMA node 0
  size mode1_ns mode2_ns ratio

---

32KiB 1.213600 1.213517 1.00x
1MiB 4.149233 6.914998 1.67x
2MiB 6.859965 23.385340 3.41x
4MiB 7.384905 31.402270 4.25x
8MiB 4.850773 34.441810 7.10x
16MiB 5.888085 35.799870 6.08x
64MiB 32.145800 36.532270 1.14x

[realmzstevenmiao]

cat smoke-mode1-vs-mode2.sh

#!/bin/bash

# SPDX-FileCopyrightText: Copyright 2026 Arm Limited and/or its affiliates
# SPDX-License-Identifier: BSD-3-Clause
#
# smoke-mode1-vs-mode2.sh
#
# Quick latency comparison of mode 1 (pair-swap shuffle) vs mode 2
# (multi-pass random chain) across a sweep of working-set sizes.
# Mode 2 defeats HW prefetchers on Neoverse V2 / Sapphire Rapids, so
# mode 2 >> mode 1 once the working set escapes L1/L2; at DRAM scale
# the ratio shrinks because mode 1's random shuffle alone already
# defeats much of the prefetcher.
#
# Usage:
#   ./smoke-mode1-vs-mode2.sh [--cpu N] [--numa-node N]
#                             [--duration SEC] [--hugepages 1G|2MB|none]

set -euo pipefail

SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
LOADED_LATENCY="${SCRIPT_DIR}/loaded-latency"

NUMA_NODE=0
CPU=""              # default: last CPU in NUMA_NODE (likely idle)
DURATION=3          # seconds per run, kept short for a smoke test
HUGEPAGE_OPT="1G"   # 1G|2MB|none — matches verify-sweep.sh default
# Probe sizes in bytes: L1, then a sweep across L2/LLC where the HW
# prefetcher gain on mode 1 is largest, then DRAM.
SIZES_BYTES=(32768 1048576 2097152 4194304 8388608 16777216 67108864)
SIZE_LABELS=("32KiB"  "1MiB"   "2MiB"   "4MiB"   "8MiB"   "16MiB"   "64MiB")

while [[ $# -gt 0 ]]; do
    case "$1" in
        --cpu)       CPU="$2";          shift 2 ;;
        --numa-node) NUMA_NODE="$2";    shift 2 ;;
        --duration)  DURATION="$2";     shift 2 ;;
        --hugepages) HUGEPAGE_OPT="$2"; shift 2 ;;
        -h|--help)
            echo "Usage: $0 [--cpu N] [--numa-node N] [--duration SEC] [--hugepages 1G|2MB|none]"
            exit 0 ;;
        *) echo "Unknown option: $1" >&2; exit 1 ;;
    esac
done

[[ -x "${LOADED_LATENCY}" ]] || { echo "Not found: ${LOADED_LATENCY} (run make)" >&2; exit 1; }
command -v numactl >/dev/null || { echo "numactl required" >&2; exit 1; }

# Default CPU = last CPU on NUMA_NODE (OS scheduler tends to leave high
# CPU IDs idle, so we get less interference without explicit isolation).
if [[ -z "${CPU}" ]]; then
    CPU=$(numactl -H | awk -v n="${NUMA_NODE}" '
        $1=="node" && $2==n && $3=="cpus:" { print $NF; exit }')
    [[ -z "${CPU}" ]] && { echo "Could not determine last CPU on NUMA node ${NUMA_NODE}" >&2; exit 1; }
fi
echo "Using CPU ${CPU} on NUMA node ${NUMA_NODE}"

run_once() {
    # Args: size_bytes mode
    local size_bytes=$1
    local mode=$2
    local cacheline_count=$((size_bytes / 64))
    (( cacheline_count < 2 )) && cacheline_count=2

    local hp_flag=""
    [[ "${HUGEPAGE_OPT}" != "none" ]] && hp_flag="--lat-use-hugepages ${HUGEPAGE_OPT}"

    local out
    out=$(numactl -N"${NUMA_NODE}" -m"${NUMA_NODE}" \
            "${LOADED_LATENCY}" -l "${CPU}" -D "${DURATION}" \
            --lat-cacheline-count "${cacheline_count}" \
            --lat-randomize-mode "${mode}" ${hp_flag} 2>&1) || true

    local lat
    lat=$(echo "$out" | awk '/^Average Latency/ {print $4; exit}')
    [[ -z "${lat}" ]] && lat="nan"
    echo "${lat}"
}

printf '%-10s %12s %12s %10s\n' "size" "mode1_ns" "mode2_ns" "ratio"
printf '%-10s %12s %12s %10s\n' "----" "--------" "--------" "-----"

for i in "${!SIZES_BYTES[@]}"; do
    sz=${SIZES_BYTES[$i]}
    label=${SIZE_LABELS[$i]}

    l1=$(run_once "$sz" 1)
    l2=$(run_once "$sz" 2)

    if [[ "$l1" == "nan" || "$l2" == "nan" ]]; then
        ratio="nan"
    else
        ratio=$(awk -v a="$l2" -v b="$l1" 'BEGIN{ if (b>0) printf "%.2f", a/b; else print "inf" }')
    fi

    printf '%-10s %12s %12s %10s\n' "${label}" "${l1}" "${l2}" "${ratio}x"
done