CPUOracle v1.0

Overview

CPUOracle is an extended dataset generation suite designed for CPU benchmarking and hardware/software Design Space Exploration (DSE). It enables users to flexibly customize datasets and concurrency (thread/process counts) for comprehensive performance evaluation.

The suite consists of two main components:

1. SPEC CPU2017 Dataset Generators: Tools to generate custom and extended datasets for all 43 workloads in the SPEC CPU2017 benchmark suite.
2. MicroBench: 10 highly customizable microbenchmarks (implemented in C and Java) with built-in dataset generators.

⚠️ IMPORTANT DISCLAIMER: This repository ONLY contains dataset generation tools. It DOES NOT include the source code or binaries for the SPEC CPU2017 benchmark suite. To use the SPEC generators, you must acquire the official SPEC CPU2017 suite from SPEC's official website.

---

Directory Structure & Features

1. MicroBench/

A collection of 10 microbenchmarks supporting flexible dataset and thread count customization.

• *.cpp / *.java: Source code for the microbenchmarks.
• run.sh: Reference script for compiling and running C workloads.
• generate_scripts.sh: Batch generator for C workloads (customize datasets and thread counts).
• generate_java_scripts.sh: Batch generator for Java workloads.
• generated_scripts/: Default output directory for generated run scripts.
• batch_run.sh: Bash script to execute all microbenchmarks sequentially.
• fast_parse_data.py: Python script to parse and aggregate the execution results.

Usage: Modify parameters inside generate_scripts.sh and generate_java_scripts.sh to generate customized execution scripts, then run them.

2. SPECCPU2017_dataset/

Contains dataset generation tools for all 43 SPEC CPU2017 workloads (Int/FP, Rate/Speed). Each directory corresponds to a specific workload. Official workload details can be found in the SPEC CPU2017 Docs.

Below is the configuration guide for customizing the datasets for each workload group:

Integer & Floating Point Workloads

• 500.perlbench_r / 600.perlbench_s:
  • Run generate-n.py. Modify number of messages to generate via runtime arguments.
• 502.gcc_r / 602.gcc_s:
  • Run generate-n.py. Accepts any successfully compilable *.c file as input. Compilation flags can be modified.
• 503.bwaves_r / 603.bwaves_s:
  • Run generate-n.py. Modify grid size parameters (nx, ny, nz) in the *.in file.
• 505.mcf_r / 605.mcf_s:
  • Run 505_generate_dataset.py. Inputs: timetabled_trip and deadhead_trip.
  • Constraint: deadhead_trip must be <= timetabled_trip * (timetabled_trip - 1) / 2.
  • Run generate-n.py to use custom_inp*.in as custom datasets.
• 507.cactuBSSN_r / 607.cactuBSSN_s:
  • Run generate-n.py. Modify 3D grid size (PUGH::local_nsize) and iterations (Cactus::cctk_itlast) in the *.par file.
• 508.namd_r:
  • Run generate-n.py. Modify the iterations parameter.
• 510.parest_r:
  • Run generate-n.py. Modify Number of experiments and Maximal number of iterations in the *.prm file.
• 511.povray_r:
  • Run generate-n.py. Modify Width and Height parameters in the *.ini file.
• 519.lbm_r / 619.lbm_s:
  • Run generate-n.py. Modify dataset size via runtime arguments (recommend changing time steps).
• 520.omnetpp_r / 620.omnetpp_s:
  • Run generate-n.py. Modify the sim-time-limit parameter in the *.ini file.
• 521.wrf_r / 621.wrf_s:
  • Run generate-n.py. Modify run_days, run_hours, run_minites, run_seconds in *.input (Note: Time range is currently limited to <= 1 day).
• 523.xalancbmk_r / 623.xalancbmk_s:
  • Run 523_generate_dataset.py (Input: Number of child nodes under the root node).
  • Run generate-n.py to select dataset_*.xml and dataset_*.xsl files.
• 525.x264_r / 625.x264_s:
  • Run generate-n.py. Modify dataset size via runtime arguments (recommend changing seek and frames).
• 526.blender_r:
  • Run generate-n.py. Modify dataset size via runtime arguments (recommend changing simulation and emulation).
• 527.cam4_r / 627.cam4_s:
  • Run generate-n.py. Modify nhtfrq in atm_in and stop_n in drv_in.
• 628.pop2_s:
  • Modify stop_n and restart_n in drv_in.in, and dt_count in pop2_in.
• 531.deepsjeng_r / 631.deepsjeng_s:
  • Run 531_generate_dataset.py. Inputs: poisons, minimum/maximum number of pieces to generate.
  • Run generate-n.py to select the generated *.txt file.
• 538.imagick_r / 638.imagick_s:
  • Run 538_generate_dataset.py. Input: Image resolution (width * height).
  • Run generate-n.py to select *.tga files.
• 541.leela_r / 641.leela_s:
  • Run 541_generate_dataset.py. Inputs: Number of games, board size, min/max moves per game.
  • Run generate-n.py to select *.sgf files.
• 544.nab_r / 644.nab_s:
  • Includes custom molecule models sourced from RCSB.
  • Run generate-n.py. Select models and random seeds via runtime arguments.
• 548.exchange2_r / 648.exchange2_s:
  • Run generate-n.py. Modify dataset size via runtime arguments.
• 549.fotonik3d_r / 649.fotonik3d_s:
  • Run generate-n.py. Modify N_x, N_y, N_z, N_t, and OBC parameters in the yee.dat file.
• 554.roms_r / 654.roms_s:
  • Run generate-n.py. Modify Lm, Mm, N, and NTIMES in *.in.x files.

  • ⚠️ CRITICAL WARNING FOR 654.roms_s: Ensure NtileI * NtileJ is multiple of the number of threads. For stability, Lm should be a multiple of NtileI, and Mm should be a multiple of NtileJ. Incorrect settings will cause runtime errors.

• 557.xz_r / 657.xz_s:
  • Run generate-n.py. Modify dataset compression levels and cache size via runtime arguments.

---

Dataset

Due to storage limitations, this repository only provides the generation tool scripts. The complete dataset is available for download at: https://resources.benchcouncil.org/datasets/CPU_Oracle.tar.gz

Citation & Reference

If you use CPUOracle in your research, please refer to our experiment and cite our paper:

• arXiv Link: https://arxiv.org/abs/2605.26643

@article{CPUOracle,
  title={Inference of Component Effect on System Performance},
  author={Wang, Chenxi and Wang, Lei and Gao, Wanling and Fan, Fanda and Kang, Guoxin and Li, Hongxiao and Su, Yuchen and Zhan, Jianfeng},
  journal={arXiv preprint arXiv:2605.26643v2},
  year={2026}
}