Newer
Older
Description and Building of the QCD Benchmark
=============================================
Description
===========
The QCD benchmark is, unlike the other benchmarks in the PRACE application benchmark suite, not a full application but a set of 7 kernels, which are representative of some of the most compute-intensive parts of lattice QCD calculations for different architectures.
The QCD benchmark suite consists of three different main application, part_cpu is based on 5 kernels of major QCD application and different major codes developed during DEISA and used since PRACE 2IP. They remain as legacy kernels while the performance tracking is shifted more towards newer computatonal kernels available in the **part-2**. The applications contains in **part-1** and **part-2** are suitable to run benchmarks on HPC machines equipped with accelerators like Nvidia GPUs or Intel Xeon Phi processors.
In the following building instruction of the CPU part are described, see for descriptions of the other parts in the different subdirectories.
Each of the 5 kernels has one test case to be used for Tier-0 and Tier-1:
**Kernel A** is derived from BQCD (Berlin Quantum ChromoDynamics program), a hybrid Monte-Carlo code that simulates Quantum Chromodynamics with dynamical standard Wilson fermions. The computations take place on a four-dimensional regular grid with periodic boundary conditions. The kernel is a standard conjugate gradient solver with even/odd pre-conditioning. Lattice size is 32x32x64x64.
**Kernel B** is derived from SU3_AHiggs, a lattice quantum chromodynamics (QCD) code intended for computing the conditions of the Early Universe. instead of "full QCD", the code applies an effective field theory, which is valid at high temperatures. In the effective theory,the lattice is 3D. Lattice size is 256x256x256.
**Kernel C** is based on benchmark kernels of the Community package openQCD, used mainly by the CLS consortium. Lattice size is 8x8x8x8. Note that Kernel C can only be run in a weak scaling mode, where each CPU stores the same local lattice size, regardless of the number of CPUs. Ideal scaling for this kernel therefore corresponds to constant execution time, and performance per peak TFlop/s is simply the reciprocal of the execution time.
**Kernel D** is based on the benchmark kernel application of tmLQCD, the community package of the Extended Twisted Mass collaboration. It consists of the core matrix-vector multiplication routine for standard Wilson fermions. The lattice size is 64x64x64x64 .
**Kernel E** consists of a full conjugate gradient solution using Wilson fermions, based on a MILC routine. The source code is also used in Part-1. The standart test cass has a lattice size of 64x64x64x64.
### Building the QCD Benchmark CPU PART in the JuBE Framework
The QCD benchmark: Part CPU is integrated in the JuBE Benchmarking Environment (www.fz-juelich.de/jsc/jube). JuBE also includes all steps to build the application.
Unpack the QCD_Source_TestCaseA.tar.gz into a directory of your choice.
After unpacking the Benchmark the following directory structure is available:
PABS/
applications/
bench/
doc/
platform/
skel/
LICENCE
The applications/ subdirectory contains the QCD benchmark applications.
The bench/ subdirectory contains the benchmark environment scripts.
The doc/ subdirectory contains the overall documentation of the framework and a tutorial.
The platform/ subdirectory holds the platform definitions as well as job submission script templates for each defined platform.
The skel/ subdirectory contains templates for analysis patterns for text output of different measurement tools.
##### Configuration
Definition files are already prepared for many platforms. If you are running on a defined platform just skip this part and go forward to QCD_Run_README.txt ("Execution").
##### The platform
A platform is defined through a set of variables in the platform.xml file, which can be found in the platform/ directory. To create a new platform entry, copy an existing platform description and modify it to fit your local setup. The variables defined here will be used by the individual applications in the later process. Best practice for the platform nomenclature would be: <vendor>-<system type>-<system name|site>. Additionally, you have to create a template batch submission script, which should be placed in a subdirectory of the platform/ directory of the same name as the platform itself. Although this nomenclature is not required by the benchmarking environment, it helps keeping track of you templates, and minimises the amount of adaptation necessary for the individual application configurations.
##### The applications
Once a platform is defined, each individual application that should be used in the benchmark (in this case the QCD application) needs to be configured for this platform. In order to configure an individual application, copy an existing top-level configuration file (e.g. prace-scaling-juqueen.xml) to the file prace-<yourplatform>.xml. Then open an editor of your choice, to adapt the file to your needs. Change the settings of the platform parameter to the name of your defined platform. The platform name can then be referenced throughout the benchmarking environment by the $platform variable.
Do the same for compile.xml, execute.xml, analyse.xml. You can find a step by step tutorial also in doc/JuBETutorial.pdf.
The compilation is part of the run of the application. Please continue with the QCD_Run_README.txt to finalize the build and to run the benchmark.