likwid

likwid (Like I knew what I am doing) is a project that contributes easy to use command-line tools for Linux to support programmers in developing high performance multi threaded programs.

It contains the following tools:

 * likwid-topology: Show the thread and cache topology
 * likwid-perfCtr: Measure hardware performance counters on Intel and AMD processors
 * likwid-features: Show and Toggle hardware prefetch control bits on Intel Core 2 processors
 * likwid-pin: Pin your threaded application without touching your code (supports pthreads, Intel OpenMP and gcc OpenMP)

There are already a bunch of performance counter tools available. Likwid stands out because:

 * No kernel patching, any vanilla kernel with the standard msr module works
 * Transparent, always clear which events are chosen, event tags have the same naming as in documentation
 * Lightweight, LIKWID on purpose does not use sampling and adds no overhead during measurement.
 * Easy to use, no java hog, simple to build, no need to touch your code, configurable from outside. Clear CLI interface
 * Multiplatform, likwid supports Intel and AMD processors
 * Up to date, likwid tries to fully support new processors as soon as possible

If you encounter problems feel free to ask questions in the User Mailing List.

NOTICE: Although Likwid is already fully usable Likwid is not yet released. I need your help to find bugs and improve Likwid to ensure a high quality for the release. Give it a try and test it. If you encounter problems write a email to me or the Forum. I promise a quick reaction time to fix errors or respond to feature requests.

What is new in this release:

• Support for Intel SandyBridge Uncore (partial) and Intel IvyBridge (only core part).
• A new application likwid-memsweeper to cleanup ccNUMA memory domains.
• Initial support for Intel Xeon Phi (KNC).
• Better support for AMD Interlagos.
• The OpenMP type is now detected automatically while pinning.
• Many bugfixes and improvements.
• The marker API works for threaded code and accessDaemon.
• Uncore support for timeline mode.
• A convenient macro wrapper for Marker API.
• Data volume as a new metric in all memory/cache groups. Updated Wiki documentation.

What is new in version 2.2.1:

• Initial support for Intel Sandy Bridge (Core) in likwid-perfctr.
• A new application, likwid-perfscope, as a frontend to the timeline mode of likwid-perfctr (experimental).
• Initial support for OpenMPI in likwid-mpirun.
• Improvements in likwid-msrD.
• A native Fortran 90 interface for the Marker API.
• A complete rewrite of the marker API.
• Simplified usage.
• Allows inclusive and overlapping regions.
• A new output filter subsystem.
• Allows you to define your own output file formats and filters.
• By default, comes with CSV and XML formats.

What is new in version 2.1.3:

• A hotfix for a build problem with direct MSR mode.
• Millisecond resolution for likwid-perfctr daemon mode.
• Improved error recovery for the MSR daemon.

What is new in version 2.1.2:

• fixes bug in likwid-mpirun script.

What is new in version 2.0:

• New application likwid-bench: A micro benchmarking platform
• likwid-pin supports now logical pinning among a node or sockets
• likwid-pin can set the NUMA mempolicy to interleave for an application
• likwid-topology also reports now on the NUMA topology
• likwid-perfCtr can print extensive documentation on the event groups
• Event groups in likwid-perfCtr can now be easily extended or modified
• Statistical output for threaded measurements in likwid-perfCtr
• Silent switch for likwid-pin
• Daemon mode for likwid-perfctr with lightweight monitoring
• likwid-perfctr supports now the Intel Nehalem EX core events
• Full support for AMD Magny Cours

What is new in version 1.1:

• Version 1.1 fixes a problem if using the perfCtr marker API together with OpenMP and using likwid-pin for pinning.

What is new in version 1.0:

• New Features:
• Lots of small improvements and fixes
• Support for Intel Westmere
• Initial support for Intel Atom
• Topology support for AMD Magny Cours