Difference between revisions of "PARSEC benchmarks"

From gem5
Jump to: navigation, search
(update information about running PARSEC on ARM)
m (Added backup mirrors for essential patch for ARM support.)
 
Line 62: Line 62:
  
 
In the following setting, 12 out of the 13 PARSEC benchmarks can be built (and run!) successfully:
 
In the following setting, 12 out of the 13 PARSEC benchmarks can be built (and run!) successfully:
* PARSEC 3.0-beta-20150206 with this [https://n.ethz.ch/~pfistchr/download/parsec-arm-static.diff patch]
+
* PARSEC 3.0-beta-20150206 with this [https://n.ethz.ch/~pfistchr/download/parsec-arm-static.diff patch] [https://mega.nz/#!Fpk03BRQ!qCKp7Tpy-OJJH7XcKfDB9o64SHR_jpRJQNA0ZqfR2Mk Mirror #1] [http://pastie.org/10844282 Mirror #2]
 
* native compilation on Debian Wheezy armhf (e.g. on qemu-system-arm -M vexpress-a9)
 
* native compilation on Debian Wheezy armhf (e.g. on qemu-system-arm -M vexpress-a9)
 
* the resulting static binaries work in gem5 (if you copy them, do not forget to copy the input files as well)
 
* the resulting static binaries work in gem5 (if you copy them, do not forget to copy the input files as well)

Latest revision as of 15:42, 19 May 2016

PARSEC v2.1

The PARSEC benchmarks have been built to run in the gem5 full-system simulation mode. The following section details references and the process for building and running the suite.

Download Pre-built M5 Disk Image

PARSEC has been built to run on gem5 with the ALPHA ISA, and disk images are available at Running PARSEC 2.1 on M5 from The University of Texas. You can download a disk image there and unzip it.

The original system files that can be downloaded from http://www.m5sim.org/Download will have the following directory structure upon extracting:

system/
       binaries/
              console
              ts_osfpal
              vmlinux
       disks/
              linux-bigswap2.img
              linux-latest.img

Now, if you are going to run for the ALPHA ISA, then you need to swap the ts_osfpal, vmlinux, linux-latest.img files in the above directory structure with the ones provided at http://www.cs.utexas.edu/~parsec_m5/

After unzipping the file, you will need to specify where gem5 should look for the disk image. In the file ./configs/common/SysPaths.py, specify the path to your disk image in the line:

path = [ ’/dist/m5/system’, ’<complete path to your disks and binaries directory>’ ]

Next, in ./configs/common/Benchmarks.py, specify the name of your disk image. For example:

return env.get(’LINUX_IMAGE’, disk(’linux-parsec-2-1-m5.img’))

You should now be set up to run with benchmarks that are on the disk image. See the section below about running gem5 for more details on how to execute the simulation.

Running PARSEC in gem5

To run PARSEC in gem5, you can specify a run script to gem5 on the command line

./build/ALPHA_FS/m5.opt ./configs/example/fs.py --script=./path/to/runscript.rcS

Where the script, runscript.rcS is a shell script that contains commands to execute the benchmark. For example:

#!/bin/sh
# File to run the blackscholes benchmark
cd /parsec/install/bin
/sbin/m5 dumpresetstats
./blackscholes 64 /parsec/install/inputs/blackscholes/in_64K.txt /parsec/install/inputs/blackscholes/prices.txt
echo "Done :D"
/sbin/m5 exit

This example changes directories to the location where the binary exists on the disk image, resets the gem5 statistics, runs the blackscholes benchmark and exits.

References

For more details on the process for building and running PARSEC for the ALPHA ISA, see Running PARSEC 2.1 on M5 from The University of Texas.

PARSEC 3.0

ARM

In the following setting, 12 out of the 13 PARSEC benchmarks can be built (and run!) successfully:

  • PARSEC 3.0-beta-20150206 with this patch Mirror #1 Mirror #2
  • native compilation on Debian Wheezy armhf (e.g. on qemu-system-arm -M vexpress-a9)
  • the resulting static binaries work in gem5 (if you copy them, do not forget to copy the input files as well)

The following benchmarks work:

  • blackscholes
  • bodytrack
  • canneal
  • dedup
  • facesim
  • ferret
  • fluidanimate
  • freqmine
  • streamcluster
  • swaptions
  • vips
  • x264

The following benchmark does not work:

  • raytrace (LRT/render.cxx uses SSE intrinsics)