Getting good performance from mdrun

Here we give an overview on the parallelization and acceleration schemes employed by GROMACS. The aim is to provide an understanding of the underlying mechanisms that make GROMACS one of the fastest molecular dynamics packages. The information presented should help choosing appropriate parallelization options, run configuration, as well as acceleration options to achieve optimal simulation performance.

The GROMACS build system and the gmx mdrun tool have a lot of built-in and configurable intelligence to detect your hardware and make pretty effective use of it. For a lot of casual and serious use of gmx mdrun, the automatic machinery works well enough. But to get the most from your hardware to maximize your scientific quality, read on!

Hardware background information

Modern computer hardware is complex and heterogeneous, so we need to discuss a little bit of background information and set up some definitions. Experienced HPC users can skip this section.

core
A hardware compute unit that actually executes instructions. There is normally more than one core in a processor, often many more.
cache
A special kind of memory local to core(s) that is much faster to access than main memory, kind of like the top of a human’s desk, compared to their filing cabinet. There are often several layers of caches associated with a core.
socket
A group of cores that share some kind of locality, such as a shared cache. This makes it more efficient to spread computational work over cores within a socket than over cores in different sockets. Modern processors often have more than one socket.
node
A group of sockets that share coarser-level locality, such as shared access to the same memory without requiring any network hardware. A normal laptop or desktop computer is a node. A node is often the smallest amount of a large compute cluster that a user can request to use.
thread
A stream of instructions for a core to execute. There are many different programming abstractions that create and manage spreading computation over multiple threads, such as OpenMP, pthreads, winthreads, CUDA, OpenCL, and OpenACC. Some kinds of hardware can map more than one software thread to a core; on Intel x86 processors this is called “hyper-threading”, while the more general concept is often called SMT for “simultaneous multi-threading”. IBM Power8 can for instance use up to 8 hardware threads per core. This feature can usually be enabled or disabled either in the hardware bios or through a setting in the Linux operating system. GROMACS can typically make use of this, for a moderate free performance boost. In most cases it will be enabled by default e.g. on new x86 processors, but in some cases the system administrators might have disabled it. If that is the case, ask if they can re-enable it for you. If you are not sure if it is enabled, check the output of the CPU information in the log file and compare with CPU specifications you find online.
thread affinity (pinning)
By default, most operating systems allow software threads to migrate between cores (or hardware threads) to help automatically balance workload. However, the performance of gmx mdrun can deteriorate if this is permitted and will degrade dramatically especially when relying on multi-threading within a rank. To avoid this, gmx mdrun will by default set the affinity of its threads to individual cores/hardware threads, unless the user or software environment has already done so (or not the entire node is used for the run, i.e. there is potential for node sharing). Setting thread affinity is sometimes called thread “pinning”.
MPI
The dominant multi-node parallelization-scheme, which provides a standardized language in which programs can be written that work across more than one node.
rank
In MPI, a rank is the smallest grouping of hardware used in the multi-node parallelization scheme. That grouping can be controlled by the user, and might correspond to a core, a socket, a node, or a group of nodes. The best choice varies with the hardware, software and compute task. Sometimes an MPI rank is called an MPI process.
GPU
A graphics processing unit, which is often faster and more efficient than conventional processors for particular kinds of compute workloads. A GPU is always associated with a particular node, and often a particular socket within that node.
OpenMP
A standardized technique supported by many compilers to share a compute workload over multiple cores. Often combined with MPI to achieve hybrid MPI/OpenMP parallelism.
CUDA
A proprietary parallel computing framework and API developed by NVIDIA that allows targeting their accelerator hardware. GROMACS uses CUDA for GPU acceleration support with NVIDIA hardware.
OpenCL
An open standard-based parallel computing framework that consists of a C99-based compiler and a programming API for targeting heterogeneous and accelerator hardware. GROMACS uses OpenCL for GPU acceleration on AMD devices (both GPUs and APUs) and Intel integrated GPUs; NVIDIA hardware is also supported.
SIMD
A type of CPU instruction by which modern CPU cores can execute multiple floating-point instructions in a single cycle.

Work distribution by parallelization in GROMACS

The algorithms in gmx mdrun and their implementations are most relevant when choosing how to make good use of the hardware. For details, see the Reference Manual. The most important of these are

Domain Decomposition
The domain decomposition (DD) algorithm decomposes the (short-ranged) component of the non-bonded interactions into domains that share spatial locality, which permits the use of efficient algorithms. Each domain handles all of the particle-particle (PP) interactions for its members, and is mapped to a single MPI rank. Within a PP rank, OpenMP threads can share the workload, and some work can be offloaded to a GPU. The PP rank also handles any bonded interactions for the members of its domain. A GPU may perform work for more than one PP rank, but it is normally most efficient to use a single PP rank per GPU and for that rank to have thousands of particles. When the work of a PP rank is done on the CPU, mdrun will make extensive use of the SIMD capabilities of the core. There are various command-line options to control the behaviour of the DD algorithm.
Particle-mesh Ewald
The particle-mesh Ewald (PME) algorithm treats the long-ranged component of the non-bonded interactions (Coulomb and possibly also Lennard-Jones). Either all, or just a subset of ranks may participate in the work for computing the long-ranged component (often inaccurately called simply the “PME” component). Because the algorithm uses a 3D FFT that requires global communication, its parallel efficiency gets worse as more ranks participate, which can mean it is fastest to use just a subset of ranks (e.g. one-quarter to one-half of the ranks). If there are separate PME ranks, then the remaining ranks handle the PP work. Otherwise, all ranks do both PP and PME work.

Parallelization schemes

GROMACS, being performance-oriented, has a strong focus on efficient parallelization. There are multiple parallelization schemes available, therefore a simulation can be run on a given hardware with different choices of run configuration.

Intra-core parallelization via SIMD: SSE, AVX, etc.

One level of performance improvement available in GROMACS is through the use of Single Instruction Multiple Data (SIMD) instructions. In detail information for those can be found under SIMD support in the installation guide.

In GROMACS, SIMD instructions are used to parallelize the parts of the code with the highest impact on performance (nonbonded and bonded force calculation, PME and neighbour searching), through the use of hardware specific SIMD kernels. Those form one of the three levels of non-bonded kernels that are available: reference or generic kernels (slow but useful for producing reference values for testing), optimized plain-C kernels (can be used cross-platform but still slow) and SIMD intrinsics accelerated kernels.

The SIMD intrinsic code is compiled by the compiler. Technically, it is possible to compile different levels of acceleration into one binary, but this is difficult to manage with acceleration in many parts of the code. Thus, you need to configure and compile GROMACS for the SIMD capabilities of the target CPU. By default, the build system will detect the highest supported acceleration of the host where the compilation is carried out. For cross-compiling for a machine with a different highest SIMD instructions set, in order to set the target acceleration, the -DGMX_SIMD CMake option can be used. To use a single installation on multiple different machines, it is convenient to compile the analysis tools with the lowest common SIMD instruction set (as these rely little on SIMD acceleration), but for best performance mdrun should be compiled be compiled separately with the highest (latest) native SIMD instruction set of the target architecture (supported by GROMACS).

Recent Intel CPU architectures bring tradeoffs between the maximum clock frequency of the CPU (ie. its speed), and the width of the SIMD instructions it executes (ie its throughput at a given speed). In particular, the Intel Skylake and Cascade Lake processors (e.g. Xeon SP Gold/Platinum), can offer better throughput when using narrower SIMD because of the better clock frequency available. Consider building mdrun configured with GMX_SIMD=AVX2_256 instead of GMX_SIMD=AVX512 for better performance in GPU accelerated or highly parallel MPI runs.

Process(-or) level parallelization via OpenMP

GROMACS mdrun supports OpenMP multithreading for all parts of the code. OpenMP is enabled by default and can be turned on/off at configure time with the GMX_OPENMP CMake variable and at run-time with the -ntomp option (or the OMP_NUM_THREADS environment variable). The OpenMP implementation is quite efficient and scales well for up to 12-24 threads on Intel and 6-8 threads on AMD CPUs.

Node level parallelization via GPU offloading and thread-MPI

Multithreading with thread-MPI

The thread-MPI library implements a subset of the MPI 1.1 specification, based on the system threading support. Both POSIX pthreads and Windows threads are supported, thus providing great portability to most UNIX/Linux and Windows operating systems. Acting as a drop-in replacement for MPI, thread-MPI enables compiling and running mdrun on a single machine (i.e. not across a network) without MPI. Additionally, it not only provides a convenient way to use computers with multicore CPU(s), but thread-MPI does in some cases make mdrun run slightly faster than with MPI.

Thread-MPI is included in the GROMACS source and it is the default parallelization since version 4.5, practically rendering the serial mdrun deprecated. Compilation with thread-MPI is controlled by the GMX_THREAD_MPI CMake variable.

Thread-MPI is compatible with most mdrun features and parallelization schemes, including OpenMP, GPUs; it is not compatible with MPI and multi-simulation runs.

By default, the thread-MPI mdrun will use all available cores in the machine by starting an appropriate number of ranks or OpenMP threads to occupy all of them. The number of ranks can be controlled using the -nt and -ntmpi options. -nt represents the total number of threads to be used (which can be a mix of thread-MPI and OpenMP threads.

Hybrid/heterogeneous acceleration

Hybrid acceleration means distributing compute work between available CPUs and GPUs to improve simulation performance. New non-bonded algorithms have been developed with the aim of efficient acceleration both on CPUs and GPUs.

The most compute-intensive parts of simulations, non-bonded force calculation, as well as possibly the PME, bonded force calculation and update and constraints can be offloaded to GPUs and carried out simultaneously with remaining CPU work. Native GPU acceleration is supported for the most commonly used algorithms in GROMACS. For more information about the GPU kernels, please see the Installation guide.

The native GPU acceleration can be turned on or off, either at run-time using the mdrun -nb option, or at configuration time using the GMX_GPU CMake variable.

To efficiently use all compute resource available, CPU and GPU computation is done simultaneously. Overlapping with the OpenMP multithreaded bonded force and PME long-range electrostatic calculations on the CPU, non-bonded forces are calculated on the GPU. Multiple GPUs, both in a single node as well as across multiple nodes, are supported using domain-decomposition. A single GPU is assigned to the non-bonded workload of a domain, therefore, the number GPUs used has to match the number of of MPI processes (or thread-MPI threads) the simulation is started with. The available CPU cores are partitioned among the processes (or thread-MPI threads) and a set of cores with a GPU do the calculations on the respective domain.

With PME electrostatics, mdrun supports automated CPU-GPU load-balancing by shifting workload from the PME mesh calculations, done on the CPU, to the particle-particle non-bonded calculations, done on the GPU. At startup a few iterations of tuning are executed during the first 100 to 1000 MD steps. These iterations involve scaling the electrostatics cut-off and PME grid spacing to determine the value that gives optimal CPU-GPU load balance. The cut-off value provided using the rcoulomb =rvdw mdp option represents the minimum electrostatics cut-off the tuning starts with and therefore should be chosen as small as possible (but still reasonable for the physics simulated). The Lennard-Jones cut-off rvdw is kept fixed. We don’t allow scaling to shorter cut-off as we don’t want to change rvdw and there would be no performance gain.

While the automated CPU-GPU load balancing always attempts to find the optimal cut-off setting, it might not always be possible to balance CPU and GPU workload. This happens when the CPU threads finish calculating the bonded forces and PME faster than the GPU the non-bonded force calculation, even with the shortest possible cut-off. In such cases the CPU will wait for the GPU and this time will show up as Wait GPU local in the cycle and timing summary table at the end of the log file.

Parallelization over multiple nodes via MPI

At the heart of the MPI parallelization in GROMACS is the neutral-territory domain decomposition with dynamic load balancing. To parallelize simulations across multiple machines (e.g. nodes of a cluster) mdrun needs to be compiled with MPI which can be enabled using the GMX_MPI CMake variable.

Controlling the domain decomposition algorithm

This section lists options that affect how the domain decomposition algorithm decomposes the workload to the available parallel hardware.

-rdd
Can be used to set the required maximum distance for inter charge-group bonded interactions. Communication for two-body bonded interactions below the non-bonded cut-off distance always comes for free with the non-bonded communication. Particles beyond the non-bonded cut-off are only communicated when they have missing bonded interactions; this means that the extra cost is minor and nearly independent of the value of -rdd. With dynamic load balancing, option -rdd also sets the lower limit for the domain decomposition cell sizes. By default -rdd is determined by gmx mdrun based on the initial coordinates. The chosen value will be a balance between interaction range and communication cost.
-ddcheck
On by default. When inter charge-group bonded interactions are beyond the bonded cut-off distance, gmx mdrun terminates with an error message. For pair interactions and tabulated bonds that do not generate exclusions, this check can be turned off with the option -noddcheck.
-rcon
When constraints are present, option -rcon influences the cell size limit as well. Particles connected by NC constraints, where NC is the LINCS order plus 1, should not be beyond the smallest cell size. A error message is generated when this happens, and the user should change the decomposition or decrease the LINCS order and increase the number of LINCS iterations. By default gmx mdrun estimates the minimum cell size required for P-LINCS in a conservative fashion. For high parallelization, it can be useful to set the distance required for P-LINCS with -rcon.
-dds
Sets the minimum allowed x, y and/or z scaling of the cells with dynamic load balancing. gmx mdrun will ensure that the cells can scale down by at least this factor. This option is used for the automated spatial decomposition (when not using -dd) as well as for determining the number of grid pulses, which in turn sets the minimum allowed cell size. Under certain circumstances the value of -dds might need to be adjusted to account for high or low spatial inhomogeneity of the system.

Multi-level parallelization: MPI and OpenMP

The multi-core trend in CPU development substantiates the need for multi-level parallelization. Current multiprocessor machines can have 2-4 CPUs with a core count as high as 64. As the memory and cache subsystem is lagging more and more behind the multicore evolution, this emphasizes non-uniform memory access (NUMA) effects, which can become a performance bottleneck. At the same time, all cores share a network interface. In a purely MPI-parallel scheme, all MPI processes use the same network interface, and although MPI intra-node communication is generally efficient, communication between nodes can become a limiting factor to parallelization. This is especially pronounced in the case of highly parallel simulations with PME (which is very communication intensive) and with ''fat'' nodes connected by a slow network. Multi-level parallelism aims to address the NUMA and communication related issues by employing efficient intra-node parallelism, typically multithreading.

Combining OpenMP with MPI creates an additional overhead especially when running separate multi-threaded PME ranks. Depending on the architecture, input system size, as well as other factors, MPI+OpenMP runs can be as fast and faster already at small number of processes (e.g. multi-processor Intel Westmere or Sandy Bridge), but can also be considerably slower (e.g. multi-processor AMD Interlagos machines). However, there is a more pronounced benefit of multi-level parallelization in highly parallel runs.

Separate PME ranks

On CPU ranks, particle-particle (PP) and PME calculations are done in the same process one after another. As PME requires all-to-all global communication, this is most of the time the limiting factor to scaling on a large number of cores. By designating a subset of ranks for PME calculations only, performance of parallel runs can be greatly improved.

OpenMP mutithreading in PME ranks is also possible. Using multi-threading in PME can can improve performance at high parallelization. The reason for this is that with N>1 threads the number of processes communicating, and therefore the number of messages, is reduced by a factor of N. But note that modern communication networks can process several messages simultaneously, such that it could be advantageous to have more processes communicating.

Separate PME ranks are not used at low parallelization, the switch at higher parallelization happens automatically (at > 16 processes). The number of PME ranks is estimated by mdrun. If the PME load is higher than the PP load, mdrun will automatically balance the load, but this leads to additional (non-bonded) calculations. This avoids the idling of a large fraction of the ranks; usually 3/4 of the ranks are PP ranks. But to ensure the best absolute performance of highly parallel runs, it is advisable to tweak this number which is automated by the tune_pme tool.

The number of PME ranks can be set manually on the mdrun command line using the -npme option, the number of PME threads can be specified on the command line with -ntomp_pme or alternatively using the GMX_PME_NUM_THREADS environment variable. The latter is especially useful when running on compute nodes with different number of cores as it enables setting different number of PME threads on different nodes.

Running mdrun within a single node

gmx mdrun can be configured and compiled in several different ways that are efficient to use within a single node. The default configuration using a suitable compiler will deploy a multi-level hybrid parallelism that uses CUDA, OpenMP and the threading platform native to the hardware. For programming convenience, in GROMACS, those native threads are used to implement on a single node the same MPI scheme as would be used between nodes, but much more efficient; this is called thread-MPI. From a user’s perspective, real MPI and thread-MPI look almost the same, and GROMACS refers to MPI ranks to mean either kind, except where noted. A real external MPI can be used for gmx mdrun within a single node, but runs more slowly than the thread-MPI version.

By default, gmx mdrun will inspect the hardware available at run time and do its best to make fairly efficient use of the whole node. The log file, stdout and stderr are used to print diagnostics that inform the user about the choices made and possible consequences.

A number of command-line parameters are available to modify the default behavior.

-nt
The total number of threads to use. The default, 0, will start as many threads as available cores. Whether the threads are thread-MPI ranks, and/or OpenMP threads within such ranks depends on other settings.
-ntmpi
The total number of thread-MPI ranks to use. The default, 0, will start one rank per GPU (if present), and otherwise one rank per core.
-ntomp
The total number of OpenMP threads per rank to start. The default, 0, will start one thread on each available core. Alternatively, mdrun will honor the appropriate system environment variable (e.g. OMP_NUM_THREADS) if set. Note that the maximum number of OpenMP threads (per rank) is, for efficiency reasons, limited to 64. While it is rarely beneficial to use a number of threads higher than this, the GMX_OPENMP_MAX_THREADS CMake variable can be used to increase the limit.
-npme
The total number of ranks to dedicate to the long-ranged component of PME, if used. The default, -1, will dedicate ranks only if the total number of threads is at least 12, and will use around a quarter of the ranks for the long-ranged component.
-ntomp_pme
When using PME with separate PME ranks, the total number of OpenMP threads per separate PME rank. The default, 0, copies the value from -ntomp.
-pin
Can be set to “auto,” “on” or “off” to control whether mdrun will attempt to set the affinity of threads to cores. Defaults to “auto,” which means that if mdrun detects that all the cores on the node are being used for mdrun, then it should behave like “on,” and attempt to set the affinities (unless they are already set by something else).
-pinoffset
If -pin on, specifies the logical core number to which mdrun should pin the first thread. When running more than one instance of mdrun on a node, use this option to to avoid pinning threads from different mdrun instances to the same core.
-pinstride
If -pin on, specifies the stride in logical core numbers for the cores to which mdrun should pin its threads. When running more than one instance of mdrun on a node, use this option to avoid pinning threads from different mdrun instances to the same core. Use the default, 0, to minimize the number of threads per physical core - this lets mdrun manage the hardware-, OS- and configuration-specific details of how to map logical cores to physical cores.
-ddorder
Can be set to “interleave,” “pp_pme” or “cartesian.” Defaults to “interleave,” which means that any separate PME ranks will be mapped to MPI ranks in an order like PP, PP, PME, PP, PP, PME, etc. This generally makes the best use of the available hardware. “pp_pme” maps all PP ranks first, then all PME ranks. “cartesian” is a special-purpose mapping generally useful only on special torus networks with accelerated global communication for Cartesian communicators. Has no effect if there are no separate PME ranks.
-nb
Used to set where to execute the short-range non-bonded interactions. Can be set to “auto”, “cpu”, “gpu.” Defaults to “auto,” which uses a compatible GPU if available. Setting “cpu” requires that no GPU is used. Setting “gpu” requires that a compatible GPU is available and will be used.
-pme
Used to set where to execute the long-range non-bonded interactions. Can be set to “auto”, “cpu”, “gpu.” Defaults to “auto,” which uses a compatible GPU if available. Setting “gpu” requires that a compatible GPU is available. Multiple PME ranks are not supported with PME on GPU, so if a GPU is used for the PME calculation -npme must be set to 1.
-bonded
Used to set where to execute the bonded interactions that are part of the PP workload for a domain. Can be set to “auto”, “cpu”, “gpu.” Defaults to “auto,” which uses a compatible CUDA GPU only when one is available, a GPU is handling short-ranged interactions, and the CPU is handling long-ranged interaction work (electrostatic or LJ). The work for the bonded interactions takes place on the same GPU as the short-ranged interactions, and cannot be independently assigned. Setting “gpu” requires that a compatible GPU is available and will be used.
-update
Used to set where to execute update and constraints, when present. Can be set to “auto”, “cpu”, “gpu.” Defaults to “auto,” which currently always uses the CPU. Setting “gpu” requires that a compatible CUDA GPU is available. Update and constraints on a GPU is currently not supported with free-energy, domain decomposition, virtual sites, Ewald surface correction, replica exchange, the pull code, orientation restraints and computational electrophysiology.
-gpu_id
A string that specifies the ID numbers of the GPUs that are available to be used by ranks on each node. For example, “12” specifies that the GPUs with IDs 1 and 2 (as reported by the GPU runtime) can be used by mdrun. This is useful when sharing a node with other computations, or if a GPU that is dedicated to a display should not be used by GROMACS. Without specifying this parameter, mdrun will utilize all GPUs. When many GPUs are present, a comma may be used to separate the IDs, so “12,13” would make GPUs 12 and 13 available to mdrun. It could be necessary to use different GPUs on different nodes of a simulation, in which case the environment variable GMX_GPU_ID can be set differently for the ranks on different nodes to achieve that result. In GROMACS versions preceding 2018 this parameter used to specify both GPU availability and GPU task assignment. The latter is now done with the -gputasks parameter.
-gputasks

A string that specifies the ID numbers of the GPUs to be used by corresponding GPU tasks on this node. For example, “0011” specifies that the first two GPU tasks will use GPU 0, and the other two use GPU 1. When using this option, the number of ranks must be known to mdrun, as well as where tasks of different types should be run, such as by using -nb gpu - only the tasks which are set to run on GPUs count for parsing the mapping. See Assigning tasks to GPUs for more details. Note that -gpu_id and -gputasks can not be used at the same time! In GROMACS versions preceding 2018 only a single type of GPU task (“PP”) could be run on any rank. Now that there is some support for running PME on GPUs, the number of GPU tasks (and the number of GPU IDs expected in the -gputasks string) can actually be 3 for a single-rank simulation. The IDs still have to be the same in this case, as using multiple GPUs per single rank is not yet implemented. The order of GPU tasks per rank in the string is PP first, PME second. The order of ranks with different kinds of GPU tasks is the same by default, but can be influenced with the -ddorder option and gets quite complex when using multiple nodes. Note that the bonded interactions for a PP task may run on the same GPU as the short-ranged work, or on the CPU, which can be controlled with the -bonded flag. The GPU task assignment (whether manually set, or automated), will be reported in the mdrun output on the first physical node of the simulation. For example:

gmx mdrun -gputasks 0001 -nb gpu -pme gpu -npme 1 -ntmpi 4

will produce the following output in the log file/terminal:

On host tcbl14 2 GPUs selected for this run.
Mapping of GPU IDs to the 4 GPU tasks in the 4 ranks on this node:
PP:0,PP:0,PP:0,PME:1

In this case, 3 ranks are set by user to compute PP work on GPU 0, and 1 rank to compute PME on GPU 1. The detailed indexing of the GPUs is also reported in the log file.

For more information about GPU tasks, please refer to Types of GPU tasks.

-pmefft
Allows choosing whether to execute the 3D FFT computation on a CPU or GPU. Can be set to “auto”, “cpu”, “gpu.”. When PME is offloaded to a GPU -pmefft gpu is the default, and the entire PME calculation is executed on the GPU. However, in some cases, e.g. with a relatively slow or older generation GPU combined with fast CPU cores in a run, moving some work off of the GPU back to the CPU by computing FFTs on the CPU can improve performance.

Examples for mdrun on one node

gmx mdrun

Starts mdrun using all the available resources. mdrun will automatically choose a fairly efficient division into thread-MPI ranks, OpenMP threads and assign work to compatible GPUs. Details will vary with hardware and the kind of simulation being run.

gmx mdrun -nt 8

Starts mdrun using 8 threads, which might be thread-MPI or OpenMP threads depending on hardware and the kind of simulation being run.

gmx mdrun -ntmpi 2 -ntomp 4

Starts mdrun using eight total threads, with two thread-MPI ranks and four OpenMP threads per rank. You should only use these options when seeking optimal performance, and must take care that the ranks you create can have all of their OpenMP threads run on the same socket. The number of ranks should be a multiple of the number of sockets, and the number of cores per node should be a multiple of the number of threads per rank.

gmx mdrun -ntmpi 4 -nb gpu -pme cpu

Starts mdrun using four thread-MPI ranks. The CPU cores available will be split evenly between the ranks using OpenMP threads. The long-range component of the forces are calculated on CPUs. This may be optimal on hardware where the CPUs are relatively powerful compared to the GPUs. The bonded part of force calculation will automatically be assigned to the GPU, since the long-range component of the forces are calculated on CPU(s).

gmx mdrun -ntmpi 1 -nb gpu -pme gpu -bonded gpu

Starts mdrun using a single thread-MPI rank that will use all available CPU cores. All interaction types that can run on a GPU will do so. This may be optimal on hardware where the CPUs are extremely weak compared to the GPUs.

gmx mdrun -ntmpi 4 -nb gpu -pme cpu -gputasks 0011

Starts mdrun using four thread-MPI ranks, and maps them to GPUs with IDs 0 and 1. The CPU cores available will be split evenly between the ranks using OpenMP threads, with the first two ranks offloading short-range nonbonded force calculations to GPU 0, and the last two ranks offloading to GPU 1. The long-range component of the forces are calculated on CPUs. This may be optimal on hardware where the CPUs are relatively powerful compared to the GPUs.

gmx mdrun -ntmpi 4 -nb gpu -pme gpu -npme 1 -gputasks 0001

Starts mdrun using four thread-MPI ranks, one of which is dedicated to the long-range PME calculation. The first 3 threads offload their short-range non-bonded calculations to the GPU with ID 0, the 4th (PME) thread offloads its calculations to the GPU with ID 1.

gmx mdrun -ntmpi 4 -nb gpu -pme gpu -npme 1 -gputasks 0011

Similar to the above example, with 3 ranks assigned to calculating short-range non-bonded forces, and one rank assigned to calculate the long-range forces. In this case, 2 of the 3 short-range ranks offload their nonbonded force calculations to GPU 0. The GPU with ID 1 calculates the short-ranged forces of the 3rd short-range rank, as well as the long-range forces of the PME-dedicated rank. Whether this or the above example is optimal will depend on the capabilities of the individual GPUs and the system composition.

gmx mdrun -gpu_id 12

Starts mdrun using GPUs with IDs 1 and 2 (e.g. because GPU 0 is dedicated to running a display). This requires two thread-MPI ranks, and will split the available CPU cores between them using OpenMP threads.

gmx mdrun -nt 6 -pin on -pinoffset 0 -pinstride 1
gmx mdrun -nt 6 -pin on -pinoffset 6 -pinstride 1

Starts two mdrun processes, each with six total threads arranged so that the processes affect each other as little as possible by being assigned to disjoint sets of physical cores. Threads will have their affinities set to particular logical cores, beginning from the first and 7th logical cores, respectively. The above would work well on an Intel CPU with six physical cores and hyper-threading enabled. Use this kind of setup only if restricting mdrun to a subset of cores to share a node with other processes. A word of caution: The mapping of logical CPUs/cores to physical cores may differ between operating systems. On Linux, cat /proc/cpuinfo can be examined to determine this mapping.

mpirun -np 2 gmx_mpi mdrun

When using an gmx mdrun compiled with external MPI, this will start two ranks and as many OpenMP threads as the hardware and MPI setup will permit. If the MPI setup is restricted to one node, then the resulting gmx mdrun will be local to that node.

Running mdrun on more than one node

This requires configuring GROMACS to build with an external MPI library. By default, this mdrun executable is run with gmx mdrun. All of the considerations for running single-node mdrun still apply, except that -ntmpi and -nt cause a fatal error, and instead the number of ranks is controlled by the MPI environment. Settings such as -npme are much more important when using multiple nodes. Configuring the MPI environment to produce one rank per core is generally good until one approaches the strong-scaling limit. At that point, using OpenMP to spread the work of an MPI rank over more than one core is needed to continue to improve absolute performance. The location of the scaling limit depends on the processor, presence of GPUs, network, and simulation algorithm, but it is worth measuring at around ~200 particles/core if you need maximum throughput.

There are further command-line parameters that are relevant in these cases.

-tunepme
Defaults to “on.” If “on,” a simulation will optimize various aspects of the PME and DD algorithms, shifting load between ranks and/or GPUs to maximize throughput. Some mdrun features are not compatible with this, and these ignore this option.
-dlb
Can be set to “auto,” “no,” or “yes.” Defaults to “auto.” Doing Dynamic Load Balancing between MPI ranks is needed to maximize performance. This is particularly important for molecular systems with heterogeneous particle or interaction density. When a certain threshold for performance loss is exceeded, DLB activates and shifts particles between ranks to improve performance. If available, using -bonded gpu is expected to improve the ability of DLB to maximize performance.

During the simulation gmx mdrun must communicate between all PP ranks to compute quantities such as kinetic energy for log file reporting, or perhaps temperature coupling. By default, this happens whenever necessary to honor several mdp options, so that the period between communication phases is the least common denominator of nstlist, nstcalcenergy, nsttcouple, and nstpcouple.

Note that -tunepme has more effect when there is more than one node, because the cost of communication for the PP and PME ranks differs. It still shifts load between PP and PME ranks, but does not change the number of separate PME ranks in use.

Note also that -dlb and -tunepme can interfere with each other, so if you experience performance variation that could result from this, you may wish to tune PME separately, and run the result with mdrun -notunepme -dlb yes.

The gmx tune_pme utility is available to search a wider range of parameter space, including making safe modifications to the tpr file, and varying -npme. It is only aware of the number of ranks created by the MPI environment, and does not explicitly manage any aspect of OpenMP during the optimization.

Examples for mdrun on more than one node

The examples and explanations for for single-node mdrun are still relevant, but -ntmpi is no longer the way to choose the number of MPI ranks.

mpirun -np 16 gmx_mpi mdrun

Starts gmx mdrun with 16 ranks, which are mapped to the hardware by the MPI library, e.g. as specified in an MPI hostfile. The available cores will be automatically split among ranks using OpenMP threads, depending on the hardware and any environment settings such as OMP_NUM_THREADS.

mpirun -np 16 gmx_mpi mdrun -npme 5

Starts gmx mdrun with 16 ranks, as above, and require that 5 of them are dedicated to the PME component.

mpirun -np 11 gmx_mpi mdrun -ntomp 2 -npme 6 -ntomp_pme 1

Starts gmx mdrun with 11 ranks, as above, and require that six of them are dedicated to the PME component with one OpenMP thread each. The remaining five do the PP component, with two OpenMP threads each.

mpirun -np 4 gmx_mpi mdrun -ntomp 6 -nb gpu -gputasks 00

Starts gmx mdrun on a machine with two nodes, using four total ranks, each rank with six OpenMP threads, and both ranks on a node sharing GPU with ID 0.

mpirun -np 8 gmx_mpi mdrun -ntomp 3 -gputasks 0000

Using a same/similar hardware as above, starts gmx mdrun on a machine with two nodes, using eight total ranks, each rank with three OpenMP threads, and all four ranks on a node sharing GPU with ID 0. This may or may not be faster than the previous setup on the same hardware.

mpirun -np 20 gmx_mpi mdrun -ntomp 4 -gputasks 00

Starts gmx mdrun with 20 ranks, and assigns the CPU cores evenly across ranks each to one OpenMP thread. This setup is likely to be suitable when there are ten nodes, each with one GPU, and each node has two sockets each of four cores.

mpirun -np 10 gmx_mpi mdrun -gpu_id 1

Starts gmx mdrun with 20 ranks, and assigns the CPU cores evenly across ranks each to one OpenMP thread. This setup is likely to be suitable when there are ten nodes, each with two GPUs, but another job on each node is using GPU 0. The job scheduler should set the affinity of threads of both jobs to their allocated cores, or the performance of mdrun will suffer greatly.

mpirun -np 20 gmx_mpi mdrun -gpu_id 01

Starts gmx mdrun with 20 ranks. This setup is likely to be suitable when there are ten nodes, each with two GPUs, but there is no need to specify -gpu_id for the normal case where all the GPUs on the node are available for use.

Approaching the scaling limit

There are several aspects of running a GROMACS simulation that are important as the number of atoms per core approaches the current scaling limit of ~100 atoms/core.

One of these is that the use of constraints = all-bonds with P-LINCS sets an artificial minimum on the size of domains. You should reconsider the use of constraints to all bonds (and bear in mind possible consequences on the safe maximum for dt), or change lincs_order and lincs_iter suitably.

Finding out how to run mdrun better

The Wallcycle module is used for runtime performance measurement of gmx mdrun. At the end of the log file of each run, the “Real cycle and time accounting” section provides a table with runtime statistics for different parts of the gmx mdrun code in rows of the table. The table contains colums indicating the number of ranks and threads that executed the respective part of the run, wall-time and cycle count aggregates (across all threads and ranks) averaged over the entire run. The last column also shows what precentage of the total runtime each row represents. Note that the gmx mdrun timer resetting functionalities (-resethway and -resetstep) reset the performance counters and therefore are useful to avoid startup overhead and performance instability (e.g. due to load balancing) at the beginning of the run.

The performance counters are:

  • Particle-particle during Particle mesh Ewald
  • Domain decomposition
  • Domain decomposition communication load
  • Domain decomposition communication bounds
  • Virtual site constraints
  • Send X to Particle mesh Ewald
  • Neighbor search
  • Launch GPU operations
  • Communication of coordinates
  • Force
  • Waiting + Communication of force
  • Particle mesh Ewald
  • PME redist. X/F
  • PME spread
  • PME gather
  • PME 3D-FFT
  • PME 3D-FFT Communication
  • PME solve Lennard-Jones
  • PME solve LJ
  • PME solve Elec
  • PME wait for particle-particle
  • Wait + Receive PME force
  • Wait GPU nonlocal
  • Wait GPU local
  • Wait PME GPU spread
  • Wait PME GPU gather
  • Reduce PME GPU Force
  • Non-bonded position/force buffer operations
  • Virtual site spread
  • COM pull force
  • AWH (accelerated weight histogram method)
  • Write trajectory
  • Update
  • Constraints
  • Communication of energies
  • Enforced rotation
  • Add rotational forces
  • Position swapping
  • Interactive MD

As performance data is collected for every run, they are essential to assessing and tuning the performance of gmx mdrun performance. Therefore, they benefit both code developers as well as users of the program. The counters are an average of the time/cycles different parts of the simulation take, hence can not directly reveal fluctuations during a single run (although comparisons across multiple runs are still very useful).

Counters will appear in an MD log file only if the related parts of the code were executed during the gmx mdrun run. There is also a special counter called “Rest” which indicates the amount of time not accounted for by any of the counters above. Therefore, a significant amount “Rest” time (more than a few percent) will often be an indication of parallelization inefficiency (e.g. serial code) and it is recommended to be reported to the developers.

An additional set of subcounters can offer more fine-grained inspection of performance. They are:

  • Domain decomposition redistribution
  • DD neighbor search grid + sort
  • DD setup communication
  • DD make topology
  • DD make constraints
  • DD topology other
  • Neighbor search grid local
  • NS grid non-local
  • NS search local
  • NS search non-local
  • Bonded force
  • Bonded-FEP force
  • Restraints force
  • Listed buffer operations
  • Nonbonded pruning
  • Nonbonded force
  • Launch non-bonded GPU tasks
  • Launch PME GPU tasks
  • Ewald force correction
  • Non-bonded position buffer operations
  • Non-bonded force buffer operations

Subcounters are geared toward developers and have to be enabled during compilation. See Build system overview for more information.

Running mdrun with GPUs

Types of GPU tasks

To better understand the later sections on different GPU use cases for calculation of short range and PME, we first introduce the concept of different GPU tasks. When thinking about running a simulation, several different kinds of interactions between the atoms have to be calculated (for more information please refer to the reference manual). The calculation can thus be split into several distinct parts that are largely independent of each other (hence can be calculated in any order, e.g. sequentially or concurrently), with the information from each of them combined at the end of time step to obtain the final forces on each atom and to propagate the system to the next time point. For a better understanding also please see the section on domain decomposition.

Of all calculations required for an MD step, GROMACS aims to optimize performance bottom-up for each step from the lowest level (SIMD unit, cores, sockets, accelerators, etc.). Therefore many of the individual computation units are highly tuned for the lowest level of hardware parallelism: the SIMD units. Additionally, with GPU accelerators used as co-processors, some of the work can be offloaded, that is calculated simultaneously/concurrently with the CPU on the accelerator device, with the result being communicated to the CPU. Right now, GROMACS supports GPU accelerator offload of two tasks: the short-range nonbonded interactions in real space, and PME.

Please note that the solving of PME on GPU is still only the initial version supporting this behaviour, and comes with a set of limitations outlined further below.

Right now, we generally support short-range nonbonded offload with and without dynamic pruning on a wide range of GPU accelerators (both NVIDIA and AMD). This is compatible with the grand majority of the features and parallelization modes and can be used to scale to large machines.

Simultaneously offloading both short-range nonbonded and long-range PME work to GPU accelerators is a new feature that that has some restrictions in terms of feature and parallelization compatibility (please see the section below).

GPU computation of short range nonbonded interactions

Using the GPU for the short-ranged nonbonded interactions provides the majority of the available speed-up compared to run using only the CPU. Here, the GPU acts as an accelerator that can effectively parallelize this problem and thus reduce the calculation time.

GPU accelerated calculation of PME

GROMACS now allows the offloading of the PME calculation to the GPU, to further reduce the load on the CPU and improve usage overlap between CPU and GPU. Here, the solving of PME will be performed in addition to the calculation of the short range interactions on the same GPU as the short range interactions.

Known limitations

Please note again the limitations outlined below!

  • PME GPU offload is supported on NVIDIA hardware with CUDA and AMD hardware with OpenCL.
  • Only a PME order of 4 is supported on GPUs.
  • PME will run on a GPU only when exactly one rank has a PME task, ie. decompositions with multiple ranks doing PME are not supported.
  • Only single precision is supported.
  • Free energy calculations where charges are perturbed are not supported, because only single PME grids can be calculated.
  • Only dynamical integrators are supported (ie. leap-frog, Velocity Verlet, stochastic dynamics)
  • LJ PME is not supported on GPUs.

GPU accelerated calculation of bonded interactions (CUDA only)

GROMACS now allows the offloading of the bonded part of the PP workload to a CUDA-compatible GPU. This is treated as part of the PP work, and requires that the short-ranged non-bonded task also runs on a GPU. It is an advantage usually only when the CPU is relatively weak compared with the GPU, perhaps because its workload is too large for the available cores. This would likely be the case for free-energy calculations.

Assigning tasks to GPUs

Depending on which tasks should be performed on which hardware, different kinds of calculations can be combined on the same or different GPUs, according to the information provided for running mdrun.

It is possible to assign the calculation of the different computational tasks to the same GPU, meaning that they will share the computational resources on the same device, or to different processing units that will each perform one task each.

One overview over the possible task assignments is given below:

GROMACS version 2018:

Two different types of assignable GPU accelerated tasks are available, NB and PME. Each PP rank has a NB task that can be offloaded to a GPU. If there is only one rank with a PME task (including if that rank is a PME-only rank), then that task can be offloaded to a GPU. Such a PME task can run wholly on the GPU, or have its latter stages run only on the CPU.

Limitations are that PME on GPU does not support PME domain decomposition, so that only one PME task can be offloaded to a single GPU assigned to a separate PME rank, while NB can be decomposed and offloaded to multiple GPUs.

GROMACS version 2019:

No new assignable GPU tasks are available, but any bonded interactions may run on the same GPU as the short-ranged interactions for a PP task. This can be influenced with the -bonded flag.

Performance considerations for GPU tasks

  1. The performance balance depends on the speed and number of CPU cores you have vs the speed and number of GPUs you have.
  2. With slow/old GPUs and/or fast/modern CPUs with many cores, it might make more sense to let the CPU do PME calculation, with the GPUs focused on the calculation of the NB.
  3. With fast/modern GPUs and/or slow/old CPUs with few cores, it generally helps to have the GPU do PME.
  4. Offloading bonded work to a GPU will often not improve simulation performance as efficient CPU-based kernels can complete the bonded computation before the GPU is done with other offloaded work. Therefore, gmx mdrun will default to no bonded offload when PME is offloaded. Typical cases where performance can be improvement with bonded offload are: with significant bonded work (e.g. pure lipid or mostly polymer systems with little solvent), with very few and/or slow CPU cores per GPU, or when the CPU does other computation (e.g. PME, free energy).
  5. It is possible to use multiple GPUs with PME offload by letting e.g. 3 MPI ranks use one GPU each for short-range interactions, while a fourth rank does the PME on its GPU.
  6. The only way to know for sure what alternative is best for your machine is to test and check performance.

Reducing overheads in GPU accelerated runs

In order for CPU cores and GPU(s) to execute concurrently, tasks are launched and executed asynchronously on the GPU(s) while the CPU cores execute non-offloaded force computation (like long-range PME electrostatics). Asynchronous task launches are handled by GPU device driver and require CPU involvement. Therefore, the work of scheduling GPU tasks will incur an overhead that can in some cases significantly delay or interfere with the CPU execution.

Delays in CPU execution are caused by the latency of launching GPU tasks, an overhead that can become significant as simulation ns/day increases (i.e. with shorter wall-time per step). The overhead is measured by gmx mdrun and reported in the performance summary section of the log file (“Launch GPU ops” row). A few percent of runtime spent in this category is normal, but in fast-iterating and multi-GPU parallel runs 10% or larger overheads can be observed. In general, a user can do little to avoid such overheads, but there are a few cases where tweaks can give performance benefits. In single-rank runs timing of GPU tasks is by default enabled and, while in most cases its impact is small, in fast runs performance can be affected. The performance impact will be most significant on NVIDIA GPUs with CUDA, less on AMD and Intel with OpenCL. In these cases, when more than a few percent of “Launch GPU ops” time is observed, it is recommended to turn off timing by setting the GMX_DISABLE_GPU_TIMING environment variable. In parallel runs with many ranks sharing a GPU, launch overheads can also be reduced by starting fewer thread-MPI or MPI ranks per GPU; e.g. most often one rank per thread or core is not optimal.

The second type of overhead, interference of the GPU driver with CPU computation, is caused by the scheduling and coordination of GPU tasks. A separate GPU driver thread can require CPU resources which may clash with the concurrently running non-offloaded tasks, potentially degrading the performance of PME or bonded force computation. This effect is most pronounced when using AMD GPUs with OpenCL with older driver releases (e.g. fglrx 12.15). To minimize the overhead it is recommended to leave a CPU hardware thread unused when launching gmx mdrun, especially on CPUs with high core counts and/or HyperThreading enabled. E.g. on a machine with a 4-core CPU and eight threads (via HyperThreading) and an AMD GPU, try gmx mdrun -ntomp 7 -pin on. This will leave free CPU resources for the GPU task scheduling reducing interference with CPU computation. Note that assigning fewer resources to gmx mdrun CPU computation involves a tradeoff which may outweigh the benefits of reduced GPU driver overhead, in particular without HyperThreading and with few CPU cores.

Running the OpenCL version of mdrun

Currently supported hardware architectures are: - GCN-based AMD GPUs; - NVIDIA GPUs (with at least OpenCL 1.2 support); - Intel iGPUs. Make sure that you have the latest drivers installed. For AMD GPUs, the compute-oriented ROCm stack is recommended; alternatively, the AMDGPU-PRO stack is also compatible; using the outdated and unsupported fglrx proprietary driver and runtime is not recommended (but for certain older hardware that may be the only way to obtain support). In addition Mesa version 17.0 or newer with LLVM 4.0 or newer is also supported. For NVIDIA GPUs, using the proprietary driver is required as the open source nouveau driver (available in Mesa) does not provide the OpenCL support. For Intel integrated GPUs, the Neo driver is recommended. TODO: add more Intel driver recommendations The minimum OpenCL version required is 1.2. See also the known limitations.

Devices from the AMD GCN architectures (all series) are compatible and regularly tested; NVIDIA Kepler and later (compute capability 3.0) are known to work, but before doing production runs always make sure that the GROMACS tests pass successfully on the hardware.

The OpenCL GPU kernels are compiled at run time. Hence, building the OpenCL program can take a few seconds, introducing a slight delay in the gmx mdrun startup. This is not normally a problem for long production MD, but you might prefer to do some kinds of work, e.g. that runs very few steps, on just the CPU (e.g. see -nb above).

The same -gpu_id option (or GMX_GPU_ID environment variable) used to select CUDA devices, or to define a mapping of GPUs to PP ranks, is used for OpenCL devices.

Some other OpenCL management environment variables may be of interest to developers.

Known limitations of the OpenCL support

Limitations in the current OpenCL support of interest to GROMACS users:

  • Intel integrated GPUs are supported. Intel CPUs and Xeon Phi are not supported.
  • Due to blocking behavior of some asynchronous task enqueuing functions in the NVIDIA OpenCL runtime, with the affected driver versions there is almost no performance gain when using NVIDIA GPUs. The issue affects NVIDIA driver versions up to 349 series, but it known to be fixed 352 and later driver releases.
  • On NVIDIA GPUs the OpenCL kernels achieve much lower performance than the equivalent CUDA kernels due to limitations of the NVIDIA OpenCL compiler.

Limitations of interest to GROMACS developers:

  • The current implementation requires a minimum execution with of 16; kernels compiled for narrower execution width (be it due to hardware requirements or compiler choice) will not be suitable and will trigger a runtime error.

Performance checklist

There are many different aspects that affect the performance of simulations in GROMACS. Most simulations require a lot of computational resources, therefore it can be worthwhile to optimize the use of those resources. Several issues mentioned in the list below could lead to a performance difference of a factor of 2. So it can be useful go through the checklist.

GROMACS configuration

  • Don’t use double precision unless you’re absolute sure you need it.
  • Compile the FFTW library (yourself) with the correct flags on x86 (in most cases, the correct flags are automatically configured).
  • On x86, use gcc or icc as the compiler (not pgi or the Cray compiler).
  • On POWER, use gcc instead of IBM’s xlc.
  • Use a new compiler version, especially for gcc (e.g. from version 5 to 6 the performance of the compiled code improved a lot).
  • MPI library: OpenMPI usually has good performance and causes little trouble.
  • Make sure your compiler supports OpenMP (some versions of Clang don’t).
  • If you have GPUs that support either CUDA or OpenCL, use them.
    • Configure with -DGMX_GPU=ON (add -DGMX_USE_OPENCL=ON for OpenCL).
    • For CUDA, use the newest CUDA available for your GPU to take advantage of the latest performance enhancements.
    • Use a recent GPU driver.
    • Make sure you use an gmx mdrun with GMX_SIMD appropriate for the CPU architecture; the log file will contain a warning note if suboptimal setting is used. However, prefer AVX2` over ``AVX512 in GPU or highly parallel MPI runs (for more information see the intra-core parallelization information).
    • If compiling on a cluster head node, make sure that GMX_SIMD is appropriate for the compute nodes.

Run setup

  • For an approximately spherical solute, use a rhombic dodecahedron unit cell.
  • When using a time-step of 2 fs, use constraints=h-bonds (and not constraints=all-bonds), since this is faster, especially with GPUs, and most force fields have been parametrized with only bonds involving hydrogens constrained.
  • You can increase the time-step to 4 or 5 fs when using virtual interaction sites (gmx pdb2gmx -vsite h).
  • For massively parallel runs with PME, you might need to try different numbers of PME ranks (gmx mdrun -npme ???) to achieve best performance; gmx tune_pme can help automate this search.
  • For massively parallel runs (also gmx mdrun -multidir), or with a slow network, global communication can become a bottleneck and you can reduce it by choosing larger periods for algorithms such as temperature and pressure coupling).

Checking and improving performance

  • Look at the end of the md.log file to see the performance and the cycle counters and wall-clock time for different parts of the MD calculation. The PP/PME load ratio is also printed, with a warning when a lot of performance is lost due to imbalance.
  • Adjust the number of PME ranks and/or the cut-off and PME grid-spacing when there is a large PP/PME imbalance. Note that even with a small reported imbalance, the automated PME-tuning might have reduced the initial imbalance. You could still gain performance by changing the mdp parameters or increasing the number of PME ranks.
  • If the neighbor searching takes a lot of time, increase nstlist. If a Verlet buffer tolerance is used, this is done automatically by gmx mdrun and the pair-list buffer is increased to keep the energy drift constant.
    • If Comm. energies takes a lot of time (a note will be printed in the log file), increase nstcalcenergy.
    • If all communication takes a lot of time, you might be running on too many cores, or you could try running combined MPI/OpenMP parallelization with 2 or 4 OpenMP threads per MPI process.