This section describes various methods for improving LAMMPS performance for different classes of problems running on different kinds of machines.5.1 Measuring performance
Before trying to make your simulation run faster, you should understand how it currently performs and where the bottlenecks are.
The best way to do this is run the your system (actual number of atoms) for a modest number of timesteps (say 100, or a few 100 at most) on several different processor counts, including a single processor if possible. Do this for an equilibrium version of your system, so that the 100-step timings are representative of a much longer run. There is typically no need to run for 1000s or timesteps to get accurate timings; you can simply extrapolate from short runs.
For the set of runs, look at the timing data printed to the screen and log file at the end of each LAMMPS run. This section of the manual has an overview.
Running on one (or a few processors) should give a good estimate of the serial performance and what portions of the timestep are taking the most time. Running the same problem on a few different processor counts should give an estimate of parallel scalability. I.e. if the simulation runs 16x faster on 16 processors, its 100% parallel efficient; if it runs 8x faster on 16 processors, it's 50% efficient.
The most important data to look at in the timing info is the timing breakdown and relative percentages. For example, trying different options for speeding up the long-range solvers will have little impact if they only consume 10% of the run time. If the pairwise time is dominating, you may want to look at GPU or OMP versions of the pair style, as discussed below. Comparing how the percentages change as you increase the processor count gives you a sense of how different operations within the timestep are scaling. Note that if you are running with a Kspace solver, there is additional output on the breakdown of the Kspace time. For PPPM, this includes the fraction spent on FFTs, which can be communication intensive.
Another important detail in the timing info are the histograms of atoms counts and neighbor counts. If these vary widely across processors, you have a load-imbalance issue. This often results in inaccurate relative timing data, because processors have to wait when communication occurs for other processors to catch up. Thus the reported times for "Communication" or "Other" may be higher than they really are, due to load-imbalance. If this is an issue, you can uncomment the MPI_Barrier() lines in src/timer.cpp, and recompile LAMMPS, to obtain synchronized timings.
NOTE: this sub-section is still a work in progress
Here is a list of general ideas for improving simulation performance. Most of them are only applicable to certain models and certain bottlenecks in the current performance, so let the timing data you intially generate be your guide. It is hard, if not impossible, to predict how much difference these options will make, since it is a function of your problem and your machine. There is no substitute for simply trying them out.
Accelerated versions of various pair_style, fixes, computes, and other commands have been added to LAMMPS, which will typically run faster than the standard non-accelerated versions, if you have the appropriate hardware on your system.
The accelerated styles have the same name as the standard styles, except that a suffix is appended. Otherwise, the syntax for the command is identical, their functionality is the same, and the numerical results it produces should also be identical, except for precision and round-off issues.
For example, all of these variants of the basic Lennard-Jones pair style exist in LAMMPS:
Assuming you have built LAMMPS with the appropriate package, these styles can be invoked by specifying them explicitly in your input script. Or you can use the -suffix command-line switch to invoke the accelerated versions automatically, without changing your input script. The suffix command allows you to set a suffix explicitly and to turn off/on the comand-line switch setting, both from within your input script.
Styles with an "opt" suffix are part of the OPT package and typically speed-up the pairwise calculations of your simulation by 5-25%.
Styles with an "omp" suffix are part of the USER-OMP package and allow a pair-style to be run in multi-threaded mode using OpenMP. This can be useful on nodes with high-core counts when using less MPI processes than cores is advantageous, e.g. when running with PPPM so that FFTs are run on fewer MPI processors or when the many MPI tasks would overload the available bandwidth for communication.
Styles with a "gpu" or "cuda" suffix are part of the GPU or USER-CUDA packages, and can be run on NVIDIA GPUs associated with your CPUs. The speed-up due to GPU usage depends on a variety of factors, as discussed below.
To see what styles are currently available in each of the accelerated packages, see Section_commands 5 of the manual. A list of accelerated styles is included in the pair, fix, compute, and kspace sections.
The following sections explain:
The final section compares and contrasts the GPU and USER-CUDA packages, since they are both designed to use NVIDIA GPU hardware.
The OPT package was developed by James Fischer (High Performance Technologies), David Richie, and Vincent Natoli (Stone Ridge Technologies). It contains a handful of pair styles whose compute() methods were rewritten in C++ templated form to reduce the overhead due to if tests and other conditional code.
The procedure for building LAMMPS with the OPT package is simple. It is the same as for any other package which has no additional library dependencies:
make yes-opt make machine
If your input script uses one of the OPT pair styles, you can run it as follows:
lmp_machine -sf opt < in.script mpirun -np 4 lmp_machine -sf opt < in.script
You should see a reduction in the "Pair time" printed out at the end of the run. On most machines and problems, this will typically be a 5 to 20% savings.
The USER-OMP package was developed by Axel Kohlmeyer at Temple University. It provides multi-threaded versions of most pair styles, all dihedral styles and a few fixes in LAMMPS. The package currently uses the OpenMP interface which requires using a specific compiler flag in the makefile to enable multiple threads; without this flag the corresponding pair styles will still be compiled and work, but do not support multi-threading.
Building LAMMPS with the USER-OMP package:
The procedure for building LAMMPS with the USER-OMP package is simple. You have to edit your machine specific makefile to add the flag to enable OpenMP support to the CCFLAGS and LINKFLAGS variables. For the GNU compilers for example this flag is called -fopenmp. Check your compiler documentation to find out which flag you need to add. The rest of the compilation is the same as for any other package which has no additional library dependencies:
make yes-user-omp make machine
Please note that this will only install accelerated versions of styles that are already installed, so you want to install this package as the last package, or else you may be missing some accelerated styles. If you plan to uninstall some package, you should first uninstall the USER-OMP package then the other package and then re-install USER-OMP, to make sure that there are no orphaned omp style files present, which would lead to compilation errors.
If your input script uses one of regular styles that are also exist as an OpenMP version in the USER-OMP package you can run it as follows:
env OMP_NUM_THREADS=4 lmp_serial -sf omp -in in.script env OMP_NUM_THREADS=2 mpirun -np 2 lmp_machine -sf omp -in in.script mpirun -x OMP_NUM_THREADS=2 -np 2 lmp_machine -sf omp -in in.script
The value of the environment variable OMP_NUM_THREADS determines how many threads per MPI task are launched. All three examples above use a total of 4 CPU cores. For different MPI implementations the method to pass the OMP_NUM_THREADS environment variable to all processes is different. Two different variants, one for MPICH and OpenMPI, respectively are shown above. Please check the documentation of your MPI installation for additional details. Alternatively, the value provided by OMP_NUM_THREADS can be overridded with the package omp command. Depending on which styles are accelerated in your input, you should see a reduction in the "Pair time" and/or "Bond time" and "Loop time" printed out at the end of the run. The optimal ratio of MPI to OpenMP can vary a lot and should always be confirmed through some benchmark runs for the current system and on the current machine.
None of the pair styles in the USER-OMP package support the "inner", "middle", "outer" options for r-RESPA integration, only the "pair" option is supported.
Parallel efficiency and performance tips:
In most simple cases the MPI parallelization in LAMMPS is more efficient than multi-threading implemented in the USER-OMP package. Also the parallel efficiency varies between individual styles. On the other hand, in many cases you still want to use the omp version - even when compiling or running without OpenMP support - since they all contain optimizations similar to those in the OPT package, which can result in serial speedup.
Using multi-threading is most effective under the following circumstances:
The best parallel efficiency from omp styles is typically achieved when there is at least one MPI task per physical processor, i.e. socket or die.
Using threads on hyper-threading enabled cores is usually counterproductive, as the cost in additional memory bandwidth requirements is not offset by the gain in CPU utilization through hyper-threading.
A description of the multi-threading strategy and some performance examples are presented here
The GPU package was developed by Mike Brown at ORNL. It provides GPU versions of several pair styles and for long-range Coulombics via the PPPM command. It has the following features:
NOTE: discuss 3 precisions if change, also have to re-link with LAMMPS always use newton off expt with differing numbers of CPUs vs GPU - can't tell what is fastest give command line switches in examples
I am not very clear to the meaning of "Max Mem / Proc" in the "GPU Time Info (average)". Is it the maximal of GPU memory used by one CPU core?
It is the maximum memory used at one time on the GPU for data storage by a single MPI process. - Mike
Hardware and software requirements:
To use this package, you currently need to have specific NVIDIA hardware and install specific NVIDIA CUDA software on your system:
Building LAMMPS with the GPU package:
As with other packages that include a separately compiled library, you need to first build the GPU library, before building LAMMPS itself. General instructions for doing this are in this section of the manual. For this package, do the following, using a Makefile in lib/gpu appropriate for your system:
cd lammps/lib/gpu make -f Makefile.linux (see further instructions in lammps/lib/gpu/README)
If you are successful, you will produce the file lib/libgpu.a.
Now you are ready to build LAMMPS with the GPU package installed:
cd lammps/src make yes-gpu make machine
Note that the lo-level Makefile (e.g. src/MAKE/Makefile.linux) has these settings: gpu_SYSINC, gpu_SYSLIB, gpu_SYSPATH. These need to be set appropriately to include the paths and settings for the CUDA system software on your machine. See src/MAKE/Makefile.g++ for an example.
When using GPUs, you are restricted to one physical GPU per LAMMPS process, which is an MPI process running on a single core or processor. Multiple MPI processes (CPU cores) can share a single GPU, and in many cases it will be more efficient to run this way.
Input script requirements:
Additional input script requirements to run pair or PPPM styles with a gpu suffix are as follows:
As an example, if you have two GPUs per node and 8 CPU cores per node, and would like to run on 4 nodes (32 cores) with dynamic balancing of force calculation across CPU and GPU cores, you could specify
package gpu force/neigh 0 1 -1
In this case, all CPU cores and GPU devices on the nodes would be utilized. Each GPU device would be shared by 4 CPU cores. The CPU cores would perform force calculations for some fraction of the particles at the same time the GPUs performed force calculation for the other particles.
As described by the package gpu command, GPU accelerated pair styles can perform computations asynchronously with CPU computations. The "Pair" time reported by LAMMPS will be the maximum of the time required to complete the CPU pair style computations and the time required to complete the GPU pair style computations. Any time spent for GPU-enabled pair styles for computations that run simultaneously with bond, angle, dihedral, improper, and long-range calculations will not be included in the "Pair" time.
When the mode setting for the package gpu command is force/neigh, the time for neighbor list calculations on the GPU will be added into the "Pair" time, not the "Neigh" time. An additional breakdown of the times required for various tasks on the GPU (data copy, neighbor calculations, force computations, etc) are output only with the LAMMPS screen output (not in the log file) at the end of each run. These timings represent total time spent on the GPU for each routine, regardless of asynchronous CPU calculations.
Generally speaking, for best performance, you should use multiple CPUs per GPU, as provided my most multi-core CPU/GPU configurations.
Because of the large number of cores within each GPU device, it may be more efficient to run on fewer processes per GPU when the number of particles per MPI process is small (100's of particles); this can be necessary to keep the GPU cores busy.
See the lammps/lib/gpu/README file for instructions on how to build the GPU library for single, mixed, or double precision. The latter requires that your GPU card support double precision.
The USER-CUDA package was developed by Christian Trott at U Technology Ilmenau in Germany. It provides NVIDIA GPU versions of many pair styles, many fixes, a few computes, and for long-range Coulombics via the PPPM command. It has the following features:
Hardware and software requirements:
To use this package, you need to have specific NVIDIA hardware and install specific NVIDIA CUDA software on your system.
Your NVIDIA GPU needs to support Compute Capability 1.3. This list may help you to find out the Compute Capability of your card:
Install the Nvidia Cuda Toolkit in version 3.2 or higher and the corresponding GPU drivers. The Nvidia Cuda SDK is not required for LAMMPSCUDA but we recommend it be installed. You can then make sure that its sample projects can be compiled without problems.
Building LAMMPS with the USER-CUDA package:
As with other packages that include a separately compiled library, you need to first build the USER-CUDA library, before building LAMMPS itself. General instructions for doing this are in this section of the manual. For this package, do the following, using settings in the lib/cuda Makefiles appropriate for your system:
precision=N to set the precision level N = 1 for single precision (default) N = 2 for double precision N = 3 for positions in double precision N = 4 for positions and velocities in double precision arch=M to set GPU compute capability M = 20 for CC2.0 (GF100/110, e.g. C2050,GTX580,GTX470) (default) M = 21 for CC2.1 (GF104/114, e.g. GTX560, GTX460, GTX450) M = 13 for CC1.3 (GF200, e.g. C1060, GTX285) prec_timer=0/1 to use hi-precision timers 0 = do not use them (default) 1 = use these timers this is usually only useful for Mac machines dbg=0/1 to activate debug mode 0 = no debug mode (default) 1 = yes debug mode this is only useful for developers cufft=1 to determine usage of CUDA FFT library 0 = no CUFFT support (default) in the future other CUDA-enabled FFT libraries might be supported
Now you are ready to build LAMMPS with the USER-CUDA package installed:
cd lammps/src make yes-user-cuda make machine
Note that the LAMMPS build references the lib/cuda/Makefile.common file to extract setting specific CUDA settings. So it is important that you have first built the cuda library (in lib/cuda) using settings appropriate to your system.
Input script requirements:
Additional input script requirements to run styles with a cuda suffix are as follows:
The USER-CUDA package offers more speed-up relative to CPU performance when the number of atoms per GPU is large, e.g. on the order of tens or hundreds of 1000s.
As noted above, this package will continue to run a simulation entirely on the GPU(s) (except for inter-processor MPI communication), for multiple timesteps, until a CPU calculation is required, either by a fix or compute that is non-GPU-ized, or until output is performed (thermo or dump snapshot or restart file). The less often this occurs, the faster your simulation will run.
Both the GPU and USER-CUDA packages accelerate a LAMMPS calculation using NVIDIA hardware, but they do it in different ways.
As a consequence, for a particular simulation on specific hardware, one package may be faster than the other. We give guidelines below, but the best way to determine which package is faster for your input script is to try both of them on your machine. See the benchmarking section below for examples where this has been done.
Guidelines for using each package optimally:
Differences between the two packages:
The LAMMPS distribution has two directories with sample input scripts for the GPU and USER-CUDA packages.
These contain input scripts for identical systems, so they can be used to benchmark the performance of both packages on your system.