Philosophy¶
This section provides reproducible benchmark results across the Dalek cluster, covering both the Dalek rack (high-performance x86 mini-PCs with discrete GPUs) and the Monolithe rack (heterogeneous SBCs).
Why Benchmark?¶
Dalek was designed from the ground up for reproducible energy-efficiency research. Benchmarks serve two goals:
- Characterize hardware: understand memory hierarchy, compute throughput, and micro-architectural behavior across a wide range of SoCs and CPUs.
- Enable fair comparisons: provide a common baseline so that results from different nodes and partitions can be meaningfully compared.
Tools¶
| Benchmark | What it measures |
|---|---|
bandwidth |
CPU memory bandwidth (read, write, copy, triad) — pinned per core |
cpufp |
CPU floating-point throughput (GFLOP/s) |
clpeak |
GPU OpenCL compute and memory bandwidth |
FMDT |
Real application: meteor detection on Full HD video |
AFF3CT |
Real application: channel coding (polar codes) |
MuRB |
N-body simulation across CPU (MIPP SIMD) and GPU (CUDA, OpenCL) |
How to Read the Curves¶
- Memory bandwidth curves show throughput (GB/s) as a function of buffer size. Three characteristic plateaus appear: L1, L2/L3, and DRAM, revealing the memory hierarchy.
- One thread per measurement: each curve corresponds to a fixed number of threads, all pinned to physical cores to avoid OS scheduling noise.
- Raw data: all CSV source files are available in
assets/raw/for independent analysis.
→ Dalek Rack Benchmarks — x86 CPUs, Nvidia/AMD/Intel GPUs
→ Monolithe Rack Benchmarks — ARM, RISC-V, Apple Silicon SBCs