3 Getting good performance from :ref:`mdrun <gmx mdrun>`
4 ======================================================
5 The |Gromacs| build system and the :ref:`gmx mdrun` tool has a lot of built-in
6 and configurable intelligence to detect your hardware and make pretty
7 effective use of that hardware. For a lot of casual and serious use of
8 :ref:`gmx mdrun`, the automatic machinery works well enough. But to get the
9 most from your hardware to maximize your scientific quality, read on!
11 Hardware background information
12 -------------------------------
13 Modern computer hardware is complex and heterogeneous, so we need to
14 discuss a little bit of background information and set up some
15 definitions. Experienced HPC users can skip this section.
20 A hardware compute unit that actually executes
21 instructions. There is normally more than one core in a
22 processor, often many more.
25 A special kind of memory local to core(s) that is much faster
26 to access than main memory, kind of like the top of a human's
27 desk, compared to their filing cabinet. There are often
28 several layers of caches associated with a core.
31 A group of cores that share some kind of locality, such as a
32 shared cache. This makes it more efficient to spread
33 computational work over cores within a socket than over cores
34 in different sockets. Modern processors often have more than
38 A group of sockets that share coarser-level locality, such as
39 shared access to the same memory without requiring any network
40 hardware. A normal laptop or desktop computer is a node. A
41 node is often the smallest amount of a large compute cluster
42 that a user can request to use.
45 A stream of instructions for a core to execute. There are many
46 different programming abstractions that create and manage
47 spreading computation over multiple threads, such as OpenMP,
48 pthreads, winthreads, CUDA, OpenCL, and OpenACC. Some kinds of
49 hardware can map more than one software thread to a core; on
50 Intel x86 processors this is called "hyper-threading", while
51 the more general concept is often called SMT for
52 "simultaneous multi-threading". IBM Power8 can for instance use
53 up to 8 hardware threads per core.
54 This feature can usually be enabled or disabled either in
55 the hardware bios or through a setting in the Linux operating
56 system. |Gromacs| can typically make use of this, for a moderate
57 free performance boost. In most cases it will be
58 enabled by default e.g. on new x86 processors, but in some cases
59 the system administrators might have disabled it. If that is the
60 case, ask if they can re-enable it for you. If you are not sure
61 if it is enabled, check the output of the CPU information in
62 the log file and compare with CPU specifications you find online.
64 thread affinity (pinning)
65 By default, most operating systems allow software threads to migrate
66 between cores (or hardware threads) to help automatically balance
67 workload. However, the performance of :ref:`gmx mdrun` can deteriorate
68 if this is permitted and will degrade dramatically especially when
69 relying on multi-threading within a rank. To avoid this,
70 :ref:`gmx mdrun` will by default
71 set the affinity of its threads to individual cores/hardware threads,
72 unless the user or software environment has already done so
73 (or not the entire node is used for the run, i.e. there is potential
75 Setting thread affinity is sometimes called thread "pinning".
78 The dominant multi-node parallelization-scheme, which provides
79 a standardized language in which programs can be written that
80 work across more than one node.
83 In MPI, a rank is the smallest grouping of hardware used in
84 the multi-node parallelization scheme. That grouping can be
85 controlled by the user, and might correspond to a core, a
86 socket, a node, or a group of nodes. The best choice varies
87 with the hardware, software and compute task. Sometimes an MPI
88 rank is called an MPI process.
91 A graphics processing unit, which is often faster and more
92 efficient than conventional processors for particular kinds of
93 compute workloads. A GPU is always associated with a
94 particular node, and often a particular socket within that
98 A standardized technique supported by many compilers to share
99 a compute workload over multiple cores. Often combined with
100 MPI to achieve hybrid MPI/OpenMP parallelism.
103 A proprietary parallel computing framework and API developed by NVIDIA
104 that allows targeting their accelerator hardware.
105 |Gromacs| uses CUDA for GPU acceleration support with NVIDIA hardware.
108 An open standard-based parallel computing framework that consists
109 of a C99-based compiler and a programming API for targeting heterogeneous
110 and accelerator hardware. |Gromacs| uses OpenCL for GPU acceleration
111 on AMD devices (both GPUs and APUs); NVIDIA hardware is also supported.
114 A type of CPU instruction by which modern CPU cores can execute large
115 numbers of floating-point instructions in a single cycle.
118 |Gromacs| background information
119 --------------------------------
120 The algorithms in :ref:`gmx mdrun` and their implementations are most relevant
121 when choosing how to make good use of the hardware. For details,
122 see the Reference Manual. The most important of these are
124 .. _gmx-domain-decomp:
129 The domain decomposition (DD) algorithm decomposes the
130 (short-ranged) component of the non-bonded interactions into
131 domains that share spatial locality, which permits the use of
132 efficient algorithms. Each domain handles all of the
133 particle-particle (PP) interactions for its members, and is
134 mapped to a single MPI rank. Within a PP rank, OpenMP threads
135 can share the workload, and some work can be off-loaded to a
136 GPU. The PP rank also handles any bonded interactions for the
137 members of its domain. A GPU may perform work for more than
138 one PP rank, but it is normally most efficient to use a single
139 PP rank per GPU and for that rank to have thousands of
140 particles. When the work of a PP rank is done on the CPU,
141 :ref:`mdrun <gmx mdrun>` will make extensive use of the SIMD
142 capabilities of the core. There are various
143 :ref:`command-line options <controlling-the-domain-decomposition-algorithm>`
144 to control the behaviour of the DD algorithm.
147 The particle-mesh Ewald (PME) algorithm treats the long-ranged
148 component of the non-bonded interactions (Coulomb and/or
149 Lennard-Jones). Either all, or just a subset of ranks may
150 participate in the work for computing the long-ranged component
151 (often inaccurately called simply the "PME"
152 component). Because the algorithm uses a 3D FFT that requires
153 global communication, its performance gets worse as more ranks
154 participate, which can mean it is fastest to use just a subset
155 of ranks (e.g. one-quarter to one-half of the ranks). If
156 there are separate PME ranks, then the remaining ranks handle
157 the PP work. Otherwise, all ranks do both PP and PME work.
159 Running :ref:`mdrun <gmx mdrun>` within a single node
160 -----------------------------------------------------
162 :ref:`gmx mdrun` can be configured and compiled in several different ways that
163 are efficient to use within a single :term:`node`. The default configuration
164 using a suitable compiler will deploy a multi-level hybrid parallelism
165 that uses CUDA, OpenMP and the threading platform native to the
166 hardware. For programming convenience, in |Gromacs|, those native
167 threads are used to implement on a single node the same MPI scheme as
168 would be used between nodes, but much more efficient; this is called
169 thread-MPI. From a user's perspective, real MPI and thread-MPI look
170 almost the same, and |Gromacs| refers to MPI ranks to mean either kind,
171 except where noted. A real external MPI can be used for :ref:`gmx mdrun` within
172 a single node, but runs more slowly than the thread-MPI version.
174 By default, :ref:`gmx mdrun` will inspect the hardware available at run time
175 and do its best to make fairly efficient use of the whole node. The
176 log file, stdout and stderr are used to print diagnostics that
177 inform the user about the choices made and possible consequences.
179 A number of command-line parameters are available to modify the default
183 The total number of threads to use. The default, 0, will start as
184 many threads as available cores. Whether the threads are
185 thread-MPI ranks, and/or OpenMP threads within such ranks depends on
189 The total number of thread-MPI ranks to use. The default, 0,
190 will start one rank per GPU (if present), and otherwise one rank
194 The total number of OpenMP threads per rank to start. The
195 default, 0, will start one thread on each available core.
196 Alternatively, :ref:`mdrun <gmx mdrun>` will honor the appropriate system
197 environment variable (e.g. ``OMP_NUM_THREADS``) if set.
200 The total number of ranks to dedicate to the long-ranged
201 component of PME, if used. The default, -1, will dedicate ranks
202 only if the total number of threads is at least 12, and will use
203 around a quarter of the ranks for the long-ranged component.
206 When using PME with separate PME ranks,
207 the total number of OpenMP threads per separate PME ranks.
208 The default, 0, copies the value from ``-ntomp``.
211 Can be set to "auto," "on" or "off" to control whether
212 :ref:`mdrun <gmx mdrun>` will attempt to set the affinity of threads to cores.
213 Defaults to "auto," which means that if :ref:`mdrun <gmx mdrun>` detects that all the
214 cores on the node are being used for :ref:`mdrun <gmx mdrun>`, then it should behave
215 like "on," and attempt to set the affinities (unless they are
216 already set by something else).
219 If ``-pin on``, specifies the logical core number to
220 which :ref:`mdrun <gmx mdrun>` should pin the first thread. When running more than
221 one instance of :ref:`mdrun <gmx mdrun>` on a node, use this option to to avoid
222 pinning threads from different :ref:`mdrun <gmx mdrun>` instances to the same core.
225 If ``-pin on``, specifies the stride in logical core
226 numbers for the cores to which :ref:`mdrun <gmx mdrun>` should pin its threads. When
227 running more than one instance of :ref:`mdrun <gmx mdrun>` on a node, use this option
228 to to avoid pinning threads from different :ref:`mdrun <gmx mdrun>` instances to the
229 same core. Use the default, 0, to minimize the number of threads
230 per physical core - this lets :ref:`mdrun <gmx mdrun>` manage the hardware-, OS- and
231 configuration-specific details of how to map logical cores to
235 Can be set to "interleave," "pp_pme" or "cartesian."
236 Defaults to "interleave," which means that any separate PME ranks
237 will be mapped to MPI ranks in an order like PP, PP, PME, PP, PP,
238 PME, ... etc. This generally makes the best use of the available
239 hardware. "pp_pme" maps all PP ranks first, then all PME
240 ranks. "cartesian" is a special-purpose mapping generally useful
241 only on special torus networks with accelerated global
242 communication for Cartesian communicators. Has no effect if there
243 are no separate PME ranks.
246 Used to set where to execute the non-bonded interactions.
247 Can be set to "auto", "cpu", "gpu."
248 Defaults to "auto," which uses a compatible GPU if available.
249 Setting "cpu" requires that no GPU is used. Setting "gpu" requires
250 that a compatible GPU be available and will be used.
253 A string that specifies the ID numbers of the GPUs that
254 are available to be used by ranks on this node. For example,
255 "12" specifies that the GPUs with IDs 1 and 2 (as reported
256 by the GPU runtime) can be used by :ref:`mdrun <gmx mdrun>`. This is useful
257 when sharing a node with other computations, or if a GPU
258 is best used to support a display. Without specifying this
259 parameter, :ref:`mdrun <gmx mdrun>` will utilize all GPUs. When many GPUs are
260 present, a comma may be used to separate the IDs, so
261 "12,13" would make GPUs 12 and 13 available to :ref:`mdrun <gmx mdrun>`.
262 It could be necessary to use different GPUs on different
263 nodes of a simulation, in which case the environment
264 variable ``GMX_GPU_ID`` can be set differently for the ranks
265 on different nodes to achieve that result.
266 In |Gromacs| versions preceding 2018 this parameter used to
267 specify both GPU availability and GPU task assignment.
268 The latter is now done with the ``-gputasks`` parameter.
271 A string that specifies the ID numbers of the GPUs to be
272 used by corresponding GPU tasks on this node. For example,
273 "0011" specifies that the first two GPU tasks will use GPU 0,
274 and the other two use GPU 1. When using this option, the
275 number of ranks must be known to :ref:`mdrun <gmx mdrun>`, as well as where
276 tasks of different types should be run, such as by using
277 ``-nb gpu`` - only the tasks which are set to run on GPUs
278 count for parsing the mapping.
279 In |Gromacs| versions preceding 2018 only a single type
280 of GPU task could be run on any rank. Now that there is some
281 support for running PME on GPUs, the number of GPU tasks
282 (and the number of GPU IDs expected in the ``-gputasks`` string)
283 can actually be 2 for a single-rank simulation. The IDs
284 still have to be the same in this case, as using multiple GPUs
285 per single rank is not yet implemented.
286 The order of GPU tasks per rank in the string is short-range first,
287 PME second. The order of ranks with different kinds of GPU tasks
288 is the same by default, but can be influenced with the ``-ddorder``
289 option and gets quite complex when using multiple nodes.
290 The GPU task assignment (whether manually set, or automated),
291 will be reported in the :ref:`mdrun <gmx mdrun>` output on
292 the first physical node of the simulation. For example:
296 gmx mdrun -gputasks 0001 -nb gpu -pme gpu -npme 1 -ntmpi 4
298 will produce the following output in the log file/terminal:
302 On host tcbl14 2 GPUs user-selected for this run.
303 Mapping of GPU IDs to the 4 GPU tasks in the 4 ranks on this node:
306 In this case, 3 ranks are set by user to compute short-range work
307 on GPU 0, and 1 rank to compute PME on GPU 1.
308 The detailed indexing of the GPUs is also reported in the log file.
310 For more information about GPU tasks, please refer to
311 :ref:`Types of GPU tasks<gmx-gpu-tasks>`.
313 Examples for :ref:`mdrun <gmx mdrun>` on one node
314 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
320 Starts :ref:`mdrun <gmx mdrun>` using all the available resources. :ref:`mdrun <gmx mdrun>`
321 will automatically choose a fairly efficient division
322 into thread-MPI ranks, OpenMP threads and assign work
323 to compatible GPUs. Details will vary with hardware
324 and the kind of simulation being run.
330 Starts :ref:`mdrun <gmx mdrun>` using 8 threads, which might be thread-MPI
331 or OpenMP threads depending on hardware and the kind
332 of simulation being run.
336 gmx mdrun -ntmpi 2 -ntomp 4
338 Starts :ref:`mdrun <gmx mdrun>` using eight total threads, with two thread-MPI
339 ranks and four OpenMP threads per rank. You should only use
340 these options when seeking optimal performance, and
341 must take care that the ranks you create can have
342 all of their OpenMP threads run on the same socket.
343 The number of ranks must be a multiple of the number of
344 sockets, and the number of cores per node must be
345 a multiple of the number of threads per rank.
351 Starts :ref:`mdrun <gmx mdrun>` using GPUs with IDs 1 and 2 (e.g. because
352 GPU 0 is dedicated to running a display). This requires
353 two thread-MPI ranks, and will split the available
354 CPU cores between them using OpenMP threads.
358 gmx mdrun -ntmpi 4 -nb gpu -gputasks 1122
360 Starts :ref:`mdrun <gmx mdrun>` using four thread-MPI ranks, and maps them
361 to GPUs with IDs 1 and 2. The CPU cores available will
362 be split evenly between the ranks using OpenMP threads.
366 gmx mdrun -nt 6 -pin on -pinoffset 0 -pinstride 1
367 gmx mdrun -nt 6 -pin on -pinoffset 6 -pinstride 1
369 Starts two :ref:`mdrun <gmx mdrun>` processes, each with six total threads
370 arranged so that the processes affect each other as little as possible by
371 being assigned to disjoint sets of physical cores.
372 Threads will have their affinities set to particular
373 logical cores, beginning from the first and 7th logical cores, respectively. The
374 above would work well on an Intel CPU with six physical cores and
375 hyper-threading enabled. Use this kind of setup only
376 if restricting :ref:`mdrun <gmx mdrun>` to a subset of cores to share a
377 node with other processes.
381 mpirun -np 2 gmx_mpi mdrun
383 When using an :ref:`gmx mdrun` compiled with external MPI,
384 this will start two ranks and as many OpenMP threads
385 as the hardware and MPI setup will permit. If the
386 MPI setup is restricted to one node, then the resulting
387 :ref:`gmx mdrun` will be local to that node.
389 Running :ref:`mdrun <gmx mdrun>` on more than one node
390 ------------------------------------------------------
391 This requires configuring |Gromacs| to build with an external MPI
392 library. By default, this :ref:`mdrun <gmx mdrun>` executable is run with
393 :ref:`mdrun_mpi`. All of the considerations for running single-node
394 :ref:`mdrun <gmx mdrun>` still apply, except that ``-ntmpi`` and ``-nt`` cause a fatal
395 error, and instead the number of ranks is controlled by the
397 Settings such as ``-npme`` are much more important when
398 using multiple nodes. Configuring the MPI environment to
399 produce one rank per core is generally good until one
400 approaches the strong-scaling limit. At that point, using
401 OpenMP to spread the work of an MPI rank over more than one
402 core is needed to continue to improve absolute performance.
403 The location of the scaling limit depends on the processor,
404 presence of GPUs, network, and simulation algorithm, but
405 it is worth measuring at around ~200 particles/core if you
406 need maximum throughput.
408 There are further command-line parameters that are relevant in these
412 Defaults to "on." If "on," a Verlet-scheme simulation will
413 optimize various aspects of the PME and DD algorithms, shifting
414 load between ranks and/or GPUs to maximize throughput. Some
415 :ref:`mdrun <gmx mdrun>` features are not compatible with this, and these ignore
419 Can be set to "auto," "no," or "yes."
420 Defaults to "auto." Doing Dynamic Load Balancing between MPI ranks
421 is needed to maximize performance. This is particularly important
422 for molecular systems with heterogeneous particle or interaction
423 density. When a certain threshold for performance loss is
424 exceeded, DLB activates and shifts particles between ranks to improve
428 During the simulation :ref:`gmx mdrun` must communicate between all ranks to
429 compute quantities such as kinetic energy. By default, this
430 happens whenever plausible, and is influenced by a lot of :ref:`[.mdp]
431 options. <mdp-general>` The period between communication phases
432 must be a multiple of :mdp:`nstlist`, and defaults to
433 the minimum of :mdp:`nstcalcenergy` and :mdp:`nstlist`.
434 ``mdrun -gcom`` sets the number of steps that must elapse between
435 such communication phases, which can improve performance when
436 running on a lot of ranks. Note that this means that _e.g._
437 temperature coupling algorithms will
438 effectively remain at constant energy until the next
439 communication phase. :ref:`gmx mdrun` will always honor the
440 setting of ``mdrun -gcom``, by changing :mdp:`nstcalcenergy`,
441 :mdp:`nstenergy`, :mdp:`nstlog`, :mdp:`nsttcouple` and/or
442 :mdp:`nstpcouple` if necessary.
444 Note that ``-tunepme`` has more effect when there is more than one
445 :term:`node`, because the cost of communication for the PP and PME
446 ranks differs. It still shifts load between PP and PME ranks, but does
447 not change the number of separate PME ranks in use.
449 Note also that ``-dlb`` and ``-tunepme`` can interfere with each other, so
450 if you experience performance variation that could result from this,
451 you may wish to tune PME separately, and run the result with ``mdrun
452 -notunepme -dlb yes``.
454 The :ref:`gmx tune_pme` utility is available to search a wider
455 range of parameter space, including making safe
456 modifications to the :ref:`tpr` file, and varying ``-npme``.
457 It is only aware of the number of ranks created by
458 the MPI environment, and does not explicitly manage
459 any aspect of OpenMP during the optimization.
461 Examples for :ref:`mdrun <gmx mdrun>` on more than one node
462 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
463 The examples and explanations for for single-node :ref:`mdrun <gmx mdrun>` are
464 still relevant, but ``-nt`` is no longer the way
465 to choose the number of MPI ranks.
469 mpirun -np 16 gmx_mpi mdrun
471 Starts :ref:`mdrun_mpi` with 16 ranks, which are mapped to
472 the hardware by the MPI library, e.g. as specified
473 in an MPI hostfile. The available cores will be
474 automatically split among ranks using OpenMP threads,
475 depending on the hardware and any environment settings
476 such as ``OMP_NUM_THREADS``.
480 mpirun -np 16 gmx_mpi mdrun -npme 5
482 Starts :ref:`mdrun_mpi` with 16 ranks, as above, and
483 require that 5 of them are dedicated to the PME
488 mpirun -np 11 gmx_mpi mdrun -ntomp 2 -npme 6 -ntomp_pme 1
490 Starts :ref:`mdrun_mpi` with 11 ranks, as above, and
491 require that six of them are dedicated to the PME
492 component with one OpenMP thread each. The remaining
493 five do the PP component, with two OpenMP threads
498 mpirun -np 4 gmx_mpi mdrun -ntomp 6 -nb gpu -gputasks 00
500 Starts :ref:`mdrun_mpi` on a machine with two nodes, using
501 four total ranks, each rank with six OpenMP threads,
502 and both ranks on a node sharing GPU with ID 0.
506 mpirun -np 8 gmx_mpi mdrun -ntomp 3 -gputasks 0000
508 Using a same/similar hardware as above,
509 starts :ref:`mdrun_mpi` on a machine with two nodes, using
510 eight total ranks, each rank with three OpenMP threads,
511 and all four ranks on a node sharing GPU with ID 0.
512 This may or may not be faster than the previous setup
513 on the same hardware.
517 mpirun -np 20 gmx_mpi mdrun -ntomp 4 -gputasks 00
519 Starts :ref:`mdrun_mpi` with 20 ranks, and assigns the CPU cores evenly
520 across ranks each to one OpenMP thread. This setup is likely to be
521 suitable when there are ten nodes, each with one GPU, and each node
522 has two sockets each of four cores.
526 mpirun -np 10 gmx_mpi mdrun -gpu_id 1
528 Starts :ref:`mdrun_mpi` with 20 ranks, and assigns the CPU cores evenly
529 across ranks each to one OpenMP thread. This setup is likely to be
530 suitable when there are ten nodes, each with two GPUs, but another
531 job on each node is using GPU 0. The job scheduler should set the
532 affinity of threads of both jobs to their allocated cores, or the
533 performance of :ref:`mdrun <gmx mdrun>` will suffer greatly.
537 mpirun -np 20 gmx_mpi mdrun -gpu_id 01
539 Starts :ref:`mdrun_mpi` with 20 ranks. This setup is likely
540 to be suitable when there are ten nodes, each with two
541 GPUs, but there is no need to specify ``-gpu_id`` for the
542 normal case where all the GPUs on the node are available
545 .. _controlling-the-domain-decomposition-algorithm:
547 Controlling the domain decomposition algorithm
548 ----------------------------------------------
549 This section lists all the options that affect how the domain
550 decomposition algorithm decomposes the workload to the available
554 Can be used to set the required maximum distance for inter
555 charge-group bonded interactions. Communication for two-body
556 bonded interactions below the non-bonded cut-off distance always
557 comes for free with the non-bonded communication. Particles beyond
558 the non-bonded cut-off are only communicated when they have
559 missing bonded interactions; this means that the extra cost is
560 minor and nearly independent of the value of ``-rdd``. With dynamic
561 load balancing, option ``-rdd`` also sets the lower limit for the
562 domain decomposition cell sizes. By default ``-rdd`` is determined
563 by :ref:`gmx mdrun` based on the initial coordinates. The chosen value will
564 be a balance between interaction range and communication cost.
567 On by default. When inter charge-group bonded interactions are
568 beyond the bonded cut-off distance, :ref:`gmx mdrun` terminates with an
569 error message. For pair interactions and tabulated bonds that do
570 not generate exclusions, this check can be turned off with the
571 option ``-noddcheck``.
574 When constraints are present, option ``-rcon`` influences
575 the cell size limit as well.
576 Particles connected by NC constraints, where NC is the LINCS order
577 plus 1, should not be beyond the smallest cell size. A error
578 message is generated when this happens, and the user should change
579 the decomposition or decrease the LINCS order and increase the
580 number of LINCS iterations. By default :ref:`gmx mdrun` estimates the
581 minimum cell size required for P-LINCS in a conservative
582 fashion. For high parallelization, it can be useful to set the
583 distance required for P-LINCS with ``-rcon``.
586 Sets the minimum allowed x, y and/or z scaling of the cells with
587 dynamic load balancing. :ref:`gmx mdrun` will ensure that the cells can
588 scale down by at least this factor. This option is used for the
589 automated spatial decomposition (when not using ``-dd``) as well as
590 for determining the number of grid pulses, which in turn sets the
591 minimum allowed cell size. Under certain circumstances the value
592 of ``-dds`` might need to be adjusted to account for high or low
593 spatial inhomogeneity of the system.
595 Finding out how to run :ref:`mdrun <gmx mdrun>` better
596 ------------------------------------------------------
598 The Wallcycle module is used for runtime performance measurement of :ref:`gmx mdrun`.
599 At the end of the log file of each run, the "Real cycle and time accounting" section
600 provides a table with runtime statistics for different parts of the :ref:`gmx mdrun` code
601 in rows of the table.
602 The table contains colums indicating the number of ranks and threads that
603 executed the respective part of the run, wall-time and cycle
604 count aggregates (across all threads and ranks) averaged over the entire run.
605 The last column also shows what precentage of the total runtime each row represents.
606 Note that the :ref:`gmx mdrun` timer resetting functionalities (`-resethway` and `-resetstep`)
607 reset the performance counters and therefore are useful to avoid startup overhead and
608 performance instability (e.g. due to load balancing) at the beginning of the run.
610 The performance counters are:
612 * Particle-particle during Particle mesh Ewald
613 * Domain decomposition
614 * Domain decomposition communication load
615 * Domain decomposition communication bounds
616 * Virtual site constraints
617 * Send X to Particle mesh Ewald
619 * Launch GPU operations
620 * Communication of coordinates
622 * Waiting + Communication of force
623 * Particle mesh Ewald
628 * PME 3D-FFT Communication
629 * PME solve Lennard-Jones
632 * PME wait for particle-particle
633 * Wait + Receive PME force
636 * Wait PME GPU spread
637 * Wait PME GPU gather
638 * Reduce PME GPU Force
639 * Non-bonded position/force buffer operations
640 * Virtual site spread
642 * AWH (accelerated weight histogram method)
646 * Communication of energies
648 * Add rotational forces
652 As performance data is collected for every run, they are essential to assessing
653 and tuning the performance of :ref:`gmx mdrun` performance. Therefore, they benefit
654 both code developers as well as users of the program.
655 The counters are an average of the time/cycles different parts of the simulation take,
656 hence can not directly reveal fluctuations during a single run (although comparisons across
657 multiple runs are still very useful).
659 Counters will appear in an MD log file only if the related parts of the code were
660 executed during the :ref:`gmx mdrun` run. There is also a special counter called "Rest" which
661 indicates the amount of time not accounted for by any of the counters above. Therefore,
662 a significant amount "Rest" time (more than a few percent) will often be an indication of
663 parallelization inefficiency (e.g. serial code) and it is recommended to be reported to the
666 An additional set of subcounters can offer more fine-grained inspection of performance. They are:
668 * Domain decomposition redistribution
669 * DD neighbor search grid + sort
670 * DD setup communication
672 * DD make constraints
674 * Neighbor search grid local
677 * NS search non-local
681 * Listed buffer operations
684 * Launch non-bonded GPU tasks
685 * Launch PME GPU tasks
686 * Ewald force correction
687 * Non-bonded position buffer operations
688 * Non-bonded force buffer operations
690 Subcounters are geared toward developers and have to be enabled during compilation. See
691 :doc:`/dev-manual/build-system` for more information.
693 TODO In future patch:
694 - red flags in log files, how to interpret wallcycle output
695 - hints to devs how to extend wallcycles
697 .. _gmx-mdrun-on-gpu:
699 Running :ref:`mdrun <gmx mdrun>` with GPUs
700 ------------------------------------------
702 NVIDIA GPUs from the professional line (Tesla or Quadro) starting with
703 the Kepler generation (compute capability 3.5 and later) support changing the
704 processor and memory clock frequency with the help of the applications clocks feature.
705 With many workloads, using higher clock rates than the default provides significant
706 performance improvements.
707 For more information see the `NVIDIA blog article`_ on this topic.
708 For |Gromacs| the highest application clock rates are optimal on all hardware
709 available to date (up to and including Maxwell, compute capability 5.2).
711 Application clocks can be set using the NVIDIA system managemet tool
712 ``nvidia-smi``. If the system permissions allow, :ref:`gmx mdrun` has
713 built-in support to set application clocks if built with :ref:`NVML support<CUDA GPU acceleration>`.
714 Note that application clocks are a global setting, hence affect the
715 performance of all applications that use the respective GPU(s).
716 For this reason, :ref:`gmx mdrun` sets application clocks at initialization
717 to the values optimal for |Gromacs| and it restores them before exiting
718 to the values found at startup, unless it detects that they were altered
721 .. _NVIDIA blog article: https://devblogs.nvidia.com/parallelforall/increase-performance-gpu-boost-k80-autoboost/
728 To better understand the later sections on different GPU use cases for
729 calculation of :ref:`short range<gmx-gpu-pp>` and :ref:`PME <gmx-gpu-pme>`,
730 we first introduce the concept of different GPU tasks. When thinking about
731 running a simulation, several different kinds of interactions between the atoms
732 have to be calculated (for more information please refer to the reference manual).
733 The calculation can thus be split into several distinct parts that are largely independent
734 of each other (hence can be calculated in any order, e.g. sequentially or concurrently),
735 with the information from each of them combined at the end of
736 time step to obtain the final forces on each atom and to propagate the system
737 to the next time point. For a better understanding also please see the section
738 on :ref:`domain decomposition <gmx-domain-decomp>`.
740 Of all calculations required for an MD step,
741 GROMACS aims to optimize performance bottom-up for each step
742 from the lowest level (SIMD unit, cores, sockets, accelerators, etc.).
743 Therefore many of the individual computation units are
744 highly tuned for the lowest level of hardware parallelism: the SIMD units.
745 Additionally, with GPU accelerators used as *co-processors*, some of the work
746 can be *offloaded*, that is calculated simultaneously/concurrently with the CPU
747 on the accelerator device, with the result being communicated to the CPU.
748 Right now, |Gromacs| supports GPU accelerator offload of two tasks:
749 the short-range :ref:`nonbonded interactions in real space <gmx-gpu-pp>`,
750 and :ref:`PME <gmx-gpu-pme>`.
752 **Please note that the solving of PME on GPU is still only the initial
753 version supporting this behaviour, and comes with a set of limitations
754 outlined further below.**
756 Right now, we generally support short-range nonbonded offload with and
757 without dynamic pruning on a wide range of GPU accelerators
758 (both NVIDIA and AMD). This is compatible with the grand majority of
759 the features and parallelization modes and can be used to scale to large machines.
761 Simultaneously offloading both short-range nonbonded and long-range
762 PME work to GPU accelerators is a new feature that that has some
763 restrictions in terms of feature and parallelization
764 compatibility (please see the :ref:`section below <gmx-pme-gpu-limitations>`).
768 GPU computation of short range nonbonded interactions
769 .....................................................
771 .. TODO make this more elaborate and include figures
773 Using the GPU for the short-ranged nonbonded interactions provides
774 the majority of the available speed-up compared to run using only the CPU.
775 Here, the GPU acts as an accelerator that can effectively parallelize
776 this problem and thus reduce the calculation time.
780 GPU accelerated calculation of PME
781 ..................................
783 .. TODO again, extend this and add some actual useful information concerning performance etc...
785 Recent additions to |Gromacs| now also allow the off-loading of the PME calculation
786 to the GPU, to further reduce the load on the CPU and improve usage overlap between
787 CPU and GPU. Here, the solving of PME will be performed in addition to the calculation
788 of the short range interactions on the same GPU as the short range interactions.
790 .. _gmx-pme-gpu-limitations:
795 **Please note again the limitations outlined below!**
797 - Only compilation with CUDA is supported.
799 - Only a PME order of 4 is supported on GPUs.
801 - PME will run on a GPU only when exactly one rank has a
802 PME task, ie. decompositions with multiple ranks doing PME are not supported.
804 - Only single precision is supported.
806 - Free energy calculations where charges are perturbed are not supported,
807 because only single PME grids can be calculated.
809 - LJ PME is not supported on GPUs.
811 Assigning tasks to GPUs
812 .......................
814 Depending on which tasks should be performed on which hardware, different kinds of
815 calculations can be combined on the same or different GPUs, according to the information
816 provided for running :ref:`mdrun <gmx mdrun>`.
818 .. Someone more knowledgeable than me should check the accuracy of this part, so that
819 I don't say something that is factually wrong :)
821 It is possible to assign the calculation of the different computational tasks to the same GPU, meaning
822 that they will share the computational resources on the same device, or to different processing units
823 that will each perform one task each.
825 One overview over the possible task assignments is given below:
827 |Gromacs| version 2018:
829 Two different types of GPU accelerated tasks are available, NB and PME.
830 Each PP rank has a NB task that can be offloaded to a GPU.
831 If there is only one rank with a PME task (including if that rank is a
832 PME-only rank), then that task can be offloaded to a GPU. Such a PME
833 task can run wholly on the GPU, or have its latter stages run only on the CPU.
835 Limitations are that PME on GPU does not support PME domain decomposition,
836 so that only one PME task can be offloaded to a single GPU
837 assigned to a separate PME rank, while NB can be decomposed and offloaded to multiple GPUs.
839 .. Future |Gromacs| versions past 2018:
841 .. Combinations of different number of NB and single PME ranks on different
842 GPUs are being planned to be implemented in the near future. In addition,
843 we plan to add support for using multiple GPUs for each rank (e.g. having one GPU
844 each to solve the NB and PME part for a single rank), and to
845 implement domain decomposition on GPUs to allow the separation of the PME
846 part to different GPU tasks.
849 Performance considerations for GPU tasks
850 ........................................
852 #) The performance balance depends on the speed and number of CPU cores you
853 have vs the speed and number of GPUs you have.
855 #) With slow/old GPUs and/or fast/modern CPUs with many
856 cores, it might make more sense to let the CPU do PME calculation,
857 with the GPUs focused on the calculation of the NB.
859 #) With fast/modern GPUs and/or slow/old CPUs with few cores,
860 it generally helps to have the GPU do PME.
862 #) It *is* possible to use multiple GPUs with PME offload
864 3 MPI ranks use one GPU each for short-range interactions,
865 while a fourth rank does the PME on its GPU.
867 #) The only way to know for sure what alternative is best for
868 your machine is to test and check performance.
870 .. TODO: we need to be more concrete here, i.e. what machine/software aspects to take into consideration, when will default run mode be using PME-GPU and when will it not, when/how should the user reason about testing different settings than the default.
872 .. TODO someone who knows about the mixed mode should comment further.
874 Reducing overheads in GPU accelerated runs
875 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
877 In order for CPU cores and GPU(s) to execute concurrently, tasks are
878 launched and executed asynchronously on the GPU(s) while the CPU cores
879 execute non-offloaded force computation (like long-range PME electrostatics).
880 Asynchronous task launches are handled by GPU device driver and
881 require CPU involvement. Therefore, the work of scheduling
882 GPU tasks will incur an overhead that can in some cases significantly
883 delay or interfere with the CPU execution.
885 Delays in CPU execution are caused by the latency of launching GPU tasks,
886 an overhead that can become significant as simulation ns/day increases
887 (i.e. with shorter wall-time per step).
888 The overhead is measured by :ref:`gmx mdrun` and reported in the performance
889 summary section of the log file ("Launch GPU ops" row).
890 A few percent of runtime spent in this category is normal,
891 but in fast-iterating and multi-GPU parallel runs 10% or larger overheads can be observed.
892 In general, a user can do little to avoid such overheads, but there
893 are a few cases where tweaks can give performance benefits.
894 In single-rank runs timing of GPU tasks is by default enabled and,
895 while in most cases its impact is small, in fast runs performance can be affected.
896 The performance impact will be most significant on NVIDIA GPUs with CUDA,
897 less on AMD with OpenCL.
898 In these cases, when more than a few percent of "Launch GPU ops" time is observed,
899 it is recommended to turn off timing by setting the ``GMX_DISABLE_GPU_TIMING``
900 environment variable.
901 In parallel runs with many ranks sharing a GPU,
902 launch overheads can also be reduced by starting fewer thread-MPI
903 or MPI ranks per GPU; e.g. most often one rank per thread or core is not optimal.
905 The second type of overhead, interference of the GPU driver with CPU computation,
906 is caused by the scheduling and coordination of GPU tasks.
907 A separate GPU driver thread can require CPU resources
908 which may clash with the concurrently running non-offloaded tasks,
909 potentially degrading the performance of PME or bonded force computation.
910 This effect is most pronounced when using AMD GPUs with OpenCL with
911 older driver releases (e.g. fglrx 12.15).
912 To minimize the overhead it is recommended to
913 leave a CPU hardware thread unused when launching :ref:`gmx mdrun`,
914 especially on CPUs with high core counts and/or HyperThreading enabled.
915 E.g. on a machine with a 4-core CPU and eight threads (via HyperThreading) and an AMD GPU,
916 try ``gmx mdrun -ntomp 7 -pin on``.
917 This will leave free CPU resources for the GPU task scheduling
918 reducing interference with CPU computation.
919 Note that assigning fewer resources to :ref:`gmx mdrun` CPU computation
920 involves a tradeoff which may outweigh the benefits of reduced GPU driver overhead,
921 in particular without HyperThreading and with few CPU cores.
923 TODO In future patch: any tips not covered above
925 Running the OpenCL version of mdrun
926 -----------------------------------
928 The current version works with GCN-based AMD GPUs, and NVIDIA CUDA
929 GPUs. Make sure that you have the latest drivers installed. For AMD GPUs,
930 the compute-oriented `ROCm <https://rocm.github.io/>`_ stack is recommended;
931 alternatively, the AMDGPU-PRO stack is also compatible; using the outdated
932 and unsupported `fglrx` proprietary driver and runtime is not recommended (but
933 for certain older hardware that may be the only way to obtain support).
934 In addition Mesa version 17.0 or newer with LLVM 4.0 or newer is also supported.
935 For NVIDIA GPUs, using the proprietary driver is
936 required as the open source nouveau driver (available in Mesa) does not
937 provide the OpenCL support.
938 The minimum OpenCL version required is |REQUIRED_OPENCL_MIN_VERSION|. See
939 also the :ref:`known limitations <opencl-known-limitations>`.
941 Devices from the AMD GCN architectures (all series) are compatible
942 and regularly tested; NVIDIA Fermi and later (compute capability 2.0)
943 are known to work, but before doing production runs always make sure that the |Gromacs| tests
944 pass successfully on the hardware.
946 The OpenCL GPU kernels are compiled at run time. Hence,
947 building the OpenCL program can take a few seconds, introducing a slight
948 delay in the :ref:`gmx mdrun` startup. This is not normally a
949 problem for long production MD, but you might prefer to do some kinds
950 of work, e.g. that runs very few steps, on just the CPU (e.g. see ``-nb`` above).
952 The same ``-gpu_id`` option (or ``GMX_GPU_ID`` environment variable)
953 used to select CUDA devices, or to define a mapping of GPUs to PP
954 ranks, is used for OpenCL devices.
956 Some other :ref:`OpenCL management <opencl-management>` environment
957 variables may be of interest to developers.
959 .. _opencl-known-limitations:
961 Known limitations of the OpenCL support
962 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
964 Limitations in the current OpenCL support of interest to |Gromacs| users:
966 - PME GPU offload is not supported with OpenCL.
967 - No Intel devices (CPUs, GPUs or Xeon Phi) are supported
968 - Due to blocking behavior of some asynchronous task enqueuing functions
969 in the NVIDIA OpenCL runtime, with the affected driver versions there is
970 almost no performance gain when using NVIDIA GPUs.
971 The issue affects NVIDIA driver versions up to 349 series, but it
972 known to be fixed 352 and later driver releases.
973 - On NVIDIA GPUs the OpenCL kernels achieve much lower performance
974 than the equivalent CUDA kernels due to limitations of the NVIDIA OpenCL
977 Limitations of interest to |Gromacs| developers:
979 - The current implementation is not compatible with OpenCL devices that are
980 not using warp/wavefronts or for which the warp/wavefront size is not a
983 Performance checklist
984 ---------------------
986 There are many different aspects that affect the performance of simulations in
987 |Gromacs|. Most simulations require a lot of computational resources, therefore
988 it can be worthwhile to optimize the use of those resources. Several issues
989 mentioned in the list below could lead to a performance difference of a factor
990 of 2. So it can be useful go through the checklist.
992 |Gromacs| configuration
993 ^^^^^^^^^^^^^^^^^^^^^^^
995 * Don't use double precision unless you're absolute sure you need it.
996 * Compile the FFTW library (yourself) with the correct flags on x86 (in most
997 cases, the correct flags are automatically configured).
998 * On x86, use gcc or icc as the compiler (not pgi or the Cray compiler).
999 * On POWER, use gcc instead of IBM's xlc.
1000 * Use a new compiler version, especially for gcc (e.g. from version 5 to 6
1001 the performance of the compiled code improved a lot).
1002 * MPI library: OpenMPI usually has good performance and causes little trouble.
1003 * Make sure your compiler supports OpenMP (some versions of Clang don't).
1004 * If you have GPUs that support either CUDA or OpenCL, use them.
1006 * Configure with ``-DGMX_GPU=ON`` (add ``-DGMX_USE_OPENCL=ON`` for OpenCL).
1007 * For CUDA, use the newest CUDA availabe for your GPU to take advantage of the
1008 latest performance enhancements.
1009 * Use a recent GPU driver.
1010 * If compiling on a cluster head node, make sure that ``GMX_SIMD``
1011 is appropriate for the compute nodes.
1016 * For an approximately spherical solute, use a rhombic dodecahedron unit cell.
1017 * When using a time-step of 2 fs, use :mdp-value:`constraints=h-bonds`
1018 (and not :mdp-value:`constraints=all-bonds`), since this is faster, especially with GPUs,
1019 and most force fields have been parametrized with only bonds involving
1020 hydrogens constrained.
1021 * You can increase the time-step to 4 or 5 fs when using virtual interaction
1022 sites (``gmx pdb2gmx -vsite h``).
1023 * For massively parallel runs with PME, you might need to try different numbers
1024 of PME ranks (``gmx mdrun -npme ???``) to achieve best performance;
1025 :ref:`gmx tune_pme` can help automate this search.
1026 * For massively parallel runs (also ``gmx mdrun -multidir``), or with a slow
1027 network, global communication can become a bottleneck and you can reduce it
1028 with ``gmx mdrun -gcom`` (note that this does affect the frequency of
1029 temperature and pressure coupling).
1031 Checking and improving performance
1032 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1034 * Look at the end of the ``md.log`` file to see the performance and the cycle
1035 counters and wall-clock time for different parts of the MD calculation. The
1036 PP/PME load ratio is also printed, with a warning when a lot of performance is
1037 lost due to imbalance.
1038 * Adjust the number of PME ranks and/or the cut-off and PME grid-spacing when
1039 there is a large PP/PME imbalance. Note that even with a small reported
1040 imbalance, the automated PME-tuning might have reduced the initial imbalance.
1041 You could still gain performance by changing the mdp parameters or increasing
1042 the number of PME ranks.
1043 * If the neighbor searching takes a lot of time, increase nstlist (with the
1044 Verlet cut-off scheme, this automatically adjusts the size of the neighbour
1045 list to do more non-bonded computation to keep energy drift constant).
1047 * If ``Comm. energies`` takes a lot of time (a note will be printed in the log
1048 file), increase nstcalcenergy or use ``mdrun -gcom``.
1049 * If all communication takes a lot of time, you might be running on too many
1050 cores, or you could try running combined MPI/OpenMP parallelization with 2
1051 or 4 OpenMP threads per MPI process.