Phiprof – Parallel Hierarchical Profiler

Copyright (c) 2011 - 2015 Finnish Meteorological Institute
Copyright (c) 2016 CSC - IT Center for Science

Licence: LGPL 3

Author: Sebastian von Alfthan (sebastian.von.alfthan at csc.fi)

Phiprof a simple library that can be used to profile parallel MPI programs. It can be used to produce a hierarchical report of time (average, max, min) spent in different timer regions. A log format can also be written, where the performance as a function of time may be reported.

Features

• Supports C++, C and Fortran 2008
• Low overhead (Less than 1 us)
• Can print the total time as a human-readable hierarchical report
• Automatically handles cases where groups of processes execute different codepaths

Not supported

• Phiprof does not automatically add regions, these have to be added manually into the code.

Installation

1. Enter src/ folder

2. Edit Makefile if needed to set the correct compilation options and compiler names. Phiprof should be compiled with a MPI compiler and openmp threading should be enabled. The compiler environment can also be set from the command line, for example "make CC=nvcc CCC=nvcc".

3. make

4. The library files are after compilation in lib/ and headers in include/ . These can manually be moved to sensible locations, or one can add the correct -I and -L flags to the compiler commands. For shared library you may also need to add the path to LD_LIBRARY_PATH

Usage

Adding timers

The C++ interface is descibed in the header file phiprof.hpp. To understand how to use it, it can also be instructive to look at the examples, in particular hello world which has more extensive comments.

Note that when adding phiprof::starts to OpenMP threaded parts of the code, one should only use the variant which uses an integer to define the region, that has been obtained by a preceeding phiprof::initializeTimer(...) call, since that is the only one that does not include a critical region synchronization.

Profiling OpenACC programs: If compiled with a PGI compiler and NVTX support, and if the OPENACC environment variable is set, each phiprof timer will also activate a corresponding NVTX region.

Running code

While running the code the phiprof reports are written out whenever the phiprof::print() function is called. On file per unique set of timers is printed, try to avoid unneccessary divergence to keep the number limited. This function can be called at any time, multiple times, and all timers do not need to be closed.

What is printed out is steered with an environment variable PHIPROF_PRINTS. It accepts a comma separated string with the following options:

• groups Prints out statistics for groups (see phiprof::initializeTimer(...) functions.
• compact Prints out timer statistics for all timers where more that 1% of time was spent
• full Prints out all timers
• detailed Prints out all timers in a alternative format with even more info on MPI (phiprof-1 style).

Default is groups,compact.