Code performance

Optimising and profiling HPC codes is key to performance, and getting the most science per kWhr.

COSMA includes many tools to help with this, and regular training activities related to these.

The SHAREing project also offers a code assessment service allowing users to submit their codes for assessment, so that potential easy performance gains can be identified.

Please fill out the form and submit your code!

The following tools are available on COSMA for profiling/benchmarking/etc.

Here is a summarized table for performance tools:

Tutorial	Description
Intel Advisor	Analyses vectorisation, threading design.
Intel VTune Profiler	Hotspot analysis, memory access patterns, threading efficiency and I/O.
AMD uProf	AMD’s profiling suite.
LIKWID	Command-line tools (`likwid-perfctr`, `likwid-topology`, `likwid-pin`, etc.) for reading hardware performance counters and controlling thread/process affinity.
gperftools	Google’s lightweight performance analysis tool.
OProfile	Sampling profiler for Linux that can profile code.
Linaro Forge (Allinea)	The DDT parallel debugger and MAP profiler for MPI/OpenMP/CUDA codes.
Intel Inspector	Dynamic analysis tool that detects memory errors and threading errors.
Valgrind	Memcheck tool finds memory errors while Cachegrind/Callgrind profile cache and branch behaviour.
MUST	Runtime correctness checker for MPI applications, catching deadlocks, datatype mismatches, resource leaks.
Score-P	Generates the profiles and OTF2 traces used by tools like Vampir and Scalasca.
Vampir	Interactive trace-visualisation tool for exploring event traces at any level of detail.
Darshan	I/O characterisation tool that captures POSIX, MPI-IO and HDF5 I/O behaviours.

Code performance

Advisor

Allinea

Darshan

gperftools

Inspector

Likwid

Must

Oprofile

Scorep

Uprof

Valgrind

Vampir

Vtune