Roofline-aware DVFS for GPUs

Date: 16-Oct-2013

Author: Cedric Nugteren (http://www.cedricnugteren.nl)

Description: This repository is an online appendix to the scientific article "Roofline-aware DVFS for GPUs"

Benchmarks

Three types of CUDA benchmarks are tested:

• Benchmarks from PolyBench/GPU
• Benchmarks from Parboil (requires Parboil datasets to be installed in ~/software/parboil-2.5/datasets/)
• Two artificial micro-benchmarks

Experimental setup

GPGPU-Sim version 3.2.1 + GPUWattch

(commit 72aaaf6b11b38121d946469f26d85315ff794f29)

Configuration for GPGPU-Sim

• Clock frequencies:

  -gpgpu_clock_domains XXX:YYY:XXX:ZZZ
  

  XXX is the halved core frequency (600-500-400-300). YYY is the full core frequency (1200-1000-800-600). ZZZ is the memory frequency (900-750-600-450).

• DRAM latencies:

  -dram_latency XXX
  

  XXX is the DRAM latency is core clock cycles, reduced when scaling the core frequency to keep the latency (in seconds) constant (100-83-76-50).

Configuration for GPUWattch

• Memory configuration:

  <param name="mc_clock" value="XXX"/>
  <param name="peak_transfer_rate" value="YYY"/>
  

  XXX is the doubled memory clock or the halved effective clock (1800-1500-1200-900). YYY is the bandwidth per memory controller (28800-24000-19200-14400).

• Clock frequencies:

  <param name="target_core_clockrate" value="XXX"/>
  <param name="clockrate" value="XXX"/>
  <param name="NOC_A" value="XXX" />
  

  XXX is either the halved or full core clock frequency in various places in the configuration settings.

• Memory power parameters:

  <param name="MEM_RD" value="XXX" />
  <param name="MEM_WR" value="YYY" />
  <param name="MEM_PRE" value="ZZZ" />
  

  XXX, YYY, and ZZZ are scaled with the core clock rate to obtain correct memory power characteristics. This has been acknowledge to be a bug in the simulator and will be repaired in the next version.

Contents of the repository

• benchmark_code

  Folder containing CUDA source code made suitable for the GPGPU-Sim simulator.

• configurations

  All the GPGPU-Sim and GPUWattch configuration files.

• results

  Folder containing the graphs as they appear in the article plus more detailed graphs. It also contains a processed database extracted from simulation data.

• simulation_data

  The raw simulation output from GPGPU-Sim and GPUWattch.

• process.r

  An R-script to process the raw simulation data and output a database in CSV format (in results folder).

• graph.r

  An R-script to generate plots based on the database generated by the process.r script.

• README

  This file.

###################################################