tf

Easily run llvm test-suite benchmarks. This is a simple test framework built using bash.

What tf can do for you?

• Run programs with PIN
• Compile and run benchmarks with your own llvm pass
• Parallel execution
• Easily collect statistics
• Run with a time limit

Benchmarks

Here is a list of benchmarks available in this repo:

• ASCI_Purple
• ASC_Sequoia
• BenchmarkGame
• BitBench
• CoyoteBench
• Dhrystone
• DOE_ProxyApps_C
• Fhourstones
• Fhourstones_31
• FreeBench
• Linpack
• llubenchmark
• mafft
• MallocBench
• McCat
• McGill
• mediabench
• MiBench
• Misc
• nbench
• NPB-serial
• Olden
• PAQ8p (C++)
• Prolangs-C
• Ptrdist
• SciMark2-C
• Shootout
• sim
• Stanford
• tramp3d-v4 (C++)
• Trimaran
• TSVC
• VersaBench
• PolyBench

Benchmarks are stored in a different repo.

Requirements

• timeout or gtimeout if you're on OS X.
• gnu-parallel
• Any version of LLVM.

Getting tf

Simply clone this repository recursively

git clone --recursive [email protected]:guilhermeleobas/tf.git

Building LLVM (if you gonna use your pass)

Below is a set of instructions to build LLVM 3.8 from source. Remember the path you build LLVM because you gonna need later.

svn co http://llvm.org/svn/llvm-project/llvm/tags/RELEASE_380/final llvm38
cd llvm38/tools
svn co http://llvm.org/svn/llvm-project/cfe/tags/RELEASE_380/final clang
cd ..
mkdir build
cd build
../configure
make -j8

Usage

The first thing you need to do is select which benchmarks you want to execute. Open benchs.sh and add the benchmark you want to run into the variable benchs.

Then, go to the file vars.sh and set the variable $LLVM_PATH to where you build LLVM. This path is something like /path/to/llvm/build/Release+Asserts/bin or /path/to/llvm/build/bin in newer versions.

Compiling Benchmarks

Simply type ./run.sh. If you only want to compile, set the flag EXEC=0 before calling run.sh.

Tip: You can skip compilation by setting COMPILE=0 before calling run.sh.

Running with a time limit

RUNTIME=8m ./run.sh or change the file vars.sh.

RUNTIME receives a number followed by an optional unit:

• s for seconds (the default)
• m for minutes
• h for hours
• d for days

After the specified time interval, timeout will send a TERM signal to the benchmark process.

Tip: Set RUNTIME=0 to run indefinitely.

Compare output

Just run with DIFF=1 and tf will compare the output produced by the binary with a reference output.

Parallel execution

We use gnu-parallel to run the benchmarks, even if you're running things sequentially. To run in parallel, change the variable $JOBS in vars.sh or call JOBS=njobs ./run.sh from the command line.

Using with Intel PIN

You need to set a few variables before. Go to the file vars.sh and change:

• PIN_PATH=/path/to/pin/
• PIN_LIB=/path/to/pintool/

The later must point to where your Pintool source code is, this way we can easily build your Pintool for you.

Now, call PIN=1 ./run.sh

Collecting stats

gnu-parallel creates a logfile called run.log. This log contains the job sequence, which host the job was run on, the start time and run time, how much data was transferred, the exit value, the signal that killed the job, and finally the command being run.

You can easily parse the logfile to a csv using python and pandas:

import pandas as pd
pd.read_csv('run.log', sep='\t').to_csv('run.csv')

You can also add your own code in the file collect.sh. tf will execute this file after all benchmarks have finished executing.

Compiling benchmarks with your own LLVM pass

See instrument.shfile. You can control how each benchmark is compiled there.

Add the path of your pass to the variable pass_path at the beginning of the instrument.sh file. Then, call tf with INSTRUMENT=1 PASS=YourPassNameHere

COMPILE=1 INSTRUMENT=1 PASS=YourPassNameHere EXEC=1 ./run.sh

---

Adding more benchmarks

1. For each folder that contains .c files, i.e., the folder that will contain the executable file that you are creating, add the following info.sh file there:

 bench_name="XX"

 source_files=( "foo.c" "bar.c" "baz.c" "..." )
 COMPILE_FLAGS=" -lm "
 COMPILER="clang"  # or clang++ for C++ programs
 RUN_OPTIONS=" irsmk_input "
 STDIN=" file.in "

The last two variables are used when tf creates the command that will be executed:

timeout -signal=TERM ${RUNTIME} ./${bench_name}.exe ${RUN_OPTIONS} < ${STDIN} > /dev/null

2. Add a function into benchs.sh, for the new benchmark.

If the benchmark does not contain subfolders, add:

function Fhourstones() {
	walk "." ;
}

otherwise, add:

function Misc() {
	dirs=($( ls -d */ )) ;   # list every folder inside Misc/
	walk "${dirs[@]}" ;
}

---

TIPS

Compiling all benchmarks takes a considerable amount of time. Is a good idea to compile them first and execute later:

COMPILE=1 EXEC=0 ./run.sh   # To compile
COMPILE=0 EXEC=1 ./run.sh   # To execute

Also, run benchmarks in parallel whenever you can. Running 220 benchmarks sequentially with a time limit of 8 minutes takes 29 hours to complete.