Causal Profiling

Causal profiling is a novel technique to measure optimization potential. This measurement matches developers' assumptions about profilers: that optimizing highly-ranked code will have the greatest impact on performance. Causal profiling measures optimization potential for serial, parallel, and asynchronous programs without instrumentation of special handling for library calls and concurrency primitives. Instead, a causal profiler uses performance experiments to predict the effect of optimizations. This allows the profiler to establish causality: "optimizing function X will have effect Y," exactly the measurement developers had assumed they were getting all along.

Requirements

Our prototype causal profiler runs on Linux and Mac OSX. The profiler relies on the LLVM infrastructure to insert profiling instrumentation. Development is done using LLVM 3.3, but other recent releases should be compatible.

Building

To build the profiler, you first need the source for LLVM 3.3.

Getting LLVM source code

Check out the LLVM source code: svn co http://llvm.org/svn/llvm-project/llvm/trunk llvm

Recommended - Check out clang:

cd llvm/tools
svn co http://llvm.org/svn/llvm-project/cfe/trunk clang
cd ../projects
svn co http://llvm/org/svn/llvm-project/compiler-rt/trunk compiler-rt
cd ../..

Getting Causal

The code for this project should be cloned into the llvm/projects directory. Starting from your original location:

cd llvm/projects
git clone http://github.com/plasma-umass/causal causal
cd ../..

Configure and build

LLVM has many different configuration options. The options below are reasonable defaults:

cd llvm
./configure --prefix=/usr/local --enable-optimized --enable-assertions --enable-shared
make install

OSX Workarounds

The causal profiler uses C++11 support, including some functionality only available in the LLVM project's libc++ runtime library on OSX. To obtain and use libc++, follow the directions available at http://libcxx.llvm.org.

Using the Profiler

The Makefile setup in tests/common.mk builds all test applications for causal profiling. To see the exact commands executed, move to an application directory under tests and run make -n.

Profiler output includes results from both slowdown and speedup experiments. Slowdown results include symbol name, file name, and line number information for each block, if available. Speedup results are in CSV format, with columns for block name, block speedup, and performance change. These results can be loaded using your favorite spreadsheet or graphing program.