Skip to content

Latest commit

 

History

History
90 lines (69 loc) · 3.37 KB

README.md

File metadata and controls

90 lines (69 loc) · 3.37 KB
Raw

Manticore Benchmarks

Below is a guide to navigating the Manticore benchmark suite.

Preparing benchmarks

Let MC equal the root directory of the Manticore compiler (e.g., <path to>/branches/swp). Let RES equal the directory in which you wish to store results generated during experiments (the default is manti-bench/results).

cd manti-bench/benchmarks
autoconf -Iconfig
./configure --with-manticore=$MC --with-resultsdir=$RES
scripts/gen-input-data.sh
scripts/gen-all-makefiles.sh

Benchmarks

All benchmark programs are stored in src/programs. Each program has its own directory. There are PML and SML versions for each program.

The PML version consists of one or more .pml files and an MLB file called mc.mlb.

The SML version consists of one or more .sml files and a smlnj.cm files and a mlton.mlb file.

The Makefile contains several build options:

- mc-par                Compiles the PML program.
- mc-seq-elision        Compiles the PML program (sequential elision, i.e., sequential binary).
- mc-par-log            Compiles the PML program (logging supported).
- mc-seq-elision-log    Compiles the PML program (sequential elision, logging supported).
- mlton                 Compiles the SML program in MLton (64 bit).
- smlnj                 Compiles the SML program in SML/NJ.
- clean                 Removes junk.

Each benchmark binary takes a single integer argument that determines the "size" of the input. The meaning of the size depends on the given benchmark. The size argument is provided by passing at the command line the argument -size .

Example: Compiling and executing 'quickhull'

Let us compile the parallel version of quickhull. cd programs/quickhull make mc-par

Let us pass quickhull 1000 random points. ./mc-par -size 1000

Executing benchmarks

To run the benchmark itself, use the driver scripts/experiment.sh. One example execution of a parallel benchmark is below. Note that each processor count ("-p N") must be specified individually. The -t flag indicates the number of iterations. The -n is the name of the results directory, and -f is the name of the file.

scripts/experiment.sh -p 1 -p 2 -p 4 -p 6 -p 8 -p 10 -p 12 -p 14 -p 16 -p 18
-p 20 -p 22 -p 24 -p 26 -p 28 -p 30 -p 32 -p 34 -p 36 -p 38 -p 40 -p 42 -p 44
-p 46 -p 48 -t 30 -n "DMM-nf" -d "DMM noflat" -f programs/dense-matrix-multiply/par-noflat

To execute a sequential benchmark, the experiment script is run in a similar fashion, though with the addition of the "-e" flag in place of the -p numbers:

scripts/experiment.sh -e -t 30 -n "DMM-seq" -d "DMM seq" -f
programs/dense-matrix-multiply/seq

Converting from JSON to CSV for plotting Results for each run will appear in the results/NAME directory and may be used as-is in their JSON format. If CSV is required (as for use in the plot-R example scripts), then use the conversion program provided in that directory:

cd results/NAME
../../plot-R/run-convert.sh NAME-date.json

TODO

Generate data automatically for smvm (input-data/mtx.txt) Convert from direct rope usage to parray syntax (once the LTS ropes are committed to SWP) programs/in-place-* are legacy and not currently working Remove some of the ICFP/worklist-specific scripts Add more docs and scripts on batch runs of benchmarks with db recording