Memory Bandwidth Characterization — Snapdragon 8 Elite (Oryon V2)

Goal: Find a per-process metric that characterizes the memory intensity of a workload (reads + writes) to estimate the contention it will cause or experience in the shared memory subsystem (LLCC and DRAM), where all processors (CPU clusters, GPU, DSP, display) compete.

Platform: Samsung Galaxy S25+, Snapdragon 8 Elite (SM8750), cpu6 (Prime cluster, 4.47 GHz).

---

Platform

Property	Value
CPU	Oryon V2 ×8: 2 Prime (cpu6–7) + 6 Gold (cpu0–5)
L2	12 MB / Prime cluster
LLCC	~112 MB (9–10 slices; confirmed by Exp 4)
DRAM	LPDDR5X, 128-bit bus
OS / kernel	Android 15, 6.6.30
Pinned core	cpu6; bwmon-llcc-prime covers cpu6–7
CPU freq	Locked to 4,473,600 kHz via /data/local/tmp/powerctl.sh

Memory Hierarchy and Measurement Points

CPU core (cpu6, Prime cluster)
  L1D: 96 KB  — private per-core
  L2:  12 MB  — shared within Prime cluster (cpu6–7)
    ↑ bus_access (PMU 0x0019): all L2 misses toward shared fabric, reads + writes
  LLCC: ~112 MB — shared by all CPU clusters, GPU, DSP, display, audio
    ↑ bwmon (bw_hwmon_meas): DRAM reads only, fires when WS > LLCC
  LPDDR5X DRAM (128-bit bus)

---

Key Findings

Use bus_access for shared-subsystem contention characterization:

• Fires at WS > L2 (> 12 MB) — captures LLCC-resident pressure that bwmon misses entirely.
• Counts reads and writes — bus_access/bwmon = 1.56 for AXPY (2 reads + 1 write).
• Per-process isolation — bwmon is cluster-level and includes background traffic.
• icc_llcc_agg is the equivalent system-level proxy (~40 ms DCVS EMA vs. ~4 ms hardware).

LLCC absorbs 30.0% ± 0.1% of all L2 misses, constant across workloads and WS sizes. Absorption depends on temporal reuse, not working-set size vs. LLCC capacity.

---

Quick Start

make NDK=/home/seonjunkim/opt/android-ndk-r29 ADB="adb -s 192.168.0.190:33291" runbench
python3 analysis/bw_bench_analysis.py

See docs/code_structure.md for all build targets, prerequisites, and detailed per-file descriptions.

---

Documentation

Document	What it covers
docs/code_structure.md	Source layout, what each file does, build instructions, critical implementation notes
docs/bandwidth_measurement.md	Memory hierarchy overview; all four metrics compared; regime table; metric selection guide
docs/pmu.md	ARM PMU architecture; bus_access event semantics; exclude_kernel=0 requirement
docs/hwmon.md	Qualcomm bw_hwmon hardware monitor; bw_hwmon_meas tracepoint
docs/icc.md	Linux ICC framework; DCVS; icc_set_bw tracepoint fields and node topology
docs/contention_metric.md	Metric decision guide: bus_access vs icc_llcc_agg vs bwmon
docs/report.md	Full experimental report: all six experiments with data tables and analysis

Suggested reading order for a new reader:

1. This README (platform + key conclusions)
2. bandwidth_measurement.md (measurement mechanisms and regimes)
3. contention_metric.md (why bus_access wins)
4. report.md (data)
5. pmu.md / hwmon.md / icc.md (mechanism deep dives)
6. code_structure.md (when working with the code)

---

Open Questions

Question	Next step
HW prefetcher disable (confirm 1.9×)	Write CPUACTLR_EL1 via kernel module
Multi-processor DRAM contention	CPU + GPU simultaneous run vs. baseline
WS=192 MB ZRAM anomaly (bwprobe)	Use mlock() to prevent page compaction
LLCC PMU access	Try Snapdragon Profiler via proprietary driver
ICC lag quantification	Longer rep windows (≥30 s) for cross-correlation