2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015,2016,2017, by the GROMACS development team, led by
5 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
6 * and including many others, as listed in the AUTHORS file in the
7 * top-level source directory and at http://www.gromacs.org.
9 * GROMACS is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public License
11 * as published by the Free Software Foundation; either version 2.1
12 * of the License, or (at your option) any later version.
14 * GROMACS is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with GROMACS; if not, see
21 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
22 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
24 * If you want to redistribute modifications to GROMACS, please
25 * consider that scientific software is very special. Version
26 * control is crucial - bugs must be traceable. We will be happy to
27 * consider code for inclusion in the official distribution, but
28 * derived work must not be called official GROMACS. Details are found
29 * in the README & COPYING files - if they are missing, get the
30 * official version at http://www.gromacs.org.
32 * To help us fund GROMACS development, we humbly ask that you cite
33 * the research papers on the package. Check out http://www.gromacs.org.
37 #include "detecthardware.h"
53 #include "thread_mpi/threads.h"
55 #include "gromacs/compat/make_unique.h"
56 #include "gromacs/gpu_utils/gpu_utils.h"
57 #include "gromacs/hardware/cpuinfo.h"
58 #include "gromacs/hardware/hardwaretopology.h"
59 #include "gromacs/hardware/hw_info.h"
60 #include "gromacs/mdtypes/commrec.h"
61 #include "gromacs/simd/support.h"
62 #include "gromacs/utility/basedefinitions.h"
63 #include "gromacs/utility/basenetwork.h"
64 #include "gromacs/utility/baseversion.h"
65 #include "gromacs/utility/cstringutil.h"
66 #include "gromacs/utility/exceptions.h"
67 #include "gromacs/utility/fatalerror.h"
68 #include "gromacs/utility/gmxassert.h"
69 #include "gromacs/utility/logger.h"
70 #include "gromacs/utility/programcontext.h"
71 #include "gromacs/utility/smalloc.h"
72 #include "gromacs/utility/stringutil.h"
73 #include "gromacs/utility/sysinfo.h"
75 #include "architecture.h"
78 # include <unistd.h> // sysconf()
84 //! Convenience macro to help us avoid ifdefs each time we use sysconf
85 #if !defined(_SC_NPROCESSORS_ONLN) && defined(_SC_NPROC_ONLN)
86 # define _SC_NPROCESSORS_ONLN _SC_NPROC_ONLN
89 //! Convenience macro to help us avoid ifdefs each time we use sysconf
90 #if !defined(_SC_NPROCESSORS_CONF) && defined(_SC_NPROC_CONF)
91 # define _SC_NPROCESSORS_CONF _SC_NPROC_CONF
94 //! Constant used to help minimize preprocessed code
95 static const bool bGPUBinary = GMX_GPU != GMX_GPU_NONE;
97 /*! \brief The hwinfo structure (common to all threads in this process).
99 * \todo This should become a shared_ptr owned by e.g. Mdrunner::runner()
100 * that is shared across any threads as needed (e.g. for thread-MPI). That
101 * offers about the same run time performance as we get here, and avoids a
102 * lot of custom code.
104 static std::unique_ptr<gmx_hw_info_t> hwinfo_g;
105 //! A reference counter for the hwinfo structure
106 static int n_hwinfo = 0;
107 //! A lock to protect the hwinfo structure
108 static tMPI_Thread_mutex_t hw_info_lock = TMPI_THREAD_MUTEX_INITIALIZER;
110 //! Detect GPUs, if that makes sense to attempt.
111 static void gmx_detect_gpus(const gmx::MDLogger &mdlog, const t_commrec *cr)
115 MPI_Comm physicalnode_comm;
117 bool isMasterRankOfNode;
119 hwinfo_g->gpu_info.bDetectGPUs =
120 (bGPUBinary && getenv("GMX_DISABLE_GPU_DETECTION") == nullptr);
121 if (!hwinfo_g->gpu_info.bDetectGPUs)
126 /* Under certain circumstances MPI ranks on the same physical node
127 * can not simultaneously access the same GPU(s). Therefore we run
128 * the detection only on one MPI rank per node and broadcast the info.
129 * Note that with thread-MPI only a single thread runs this code.
131 * NOTE: We can't broadcast gpu_info with OpenCL as the device and platform
132 * ID stored in the structure are unique for each rank (even if a device
133 * is shared by multiple ranks).
135 * TODO: We should also do CPU hardware detection only once on each
136 * physical node and broadcast it, instead of do it on every MPI rank.
139 /* A split of MPI_COMM_WORLD over physical nodes is only required here,
140 * so we create and destroy it locally.
142 MPI_Comm_rank(MPI_COMM_WORLD, &rank_world);
143 MPI_Comm_split(MPI_COMM_WORLD, gmx_physicalnode_id_hash(),
144 rank_world, &physicalnode_comm);
147 MPI_Comm_rank(physicalnode_comm, &rankOnNode);
148 isMasterRankOfNode = (rankOnNode == 0);
150 GMX_UNUSED_VALUE(cr);
152 // Here there should be only one process, because if we are using
153 // thread-MPI, only one thread is active so far. So we check this.
154 GMX_RELEASE_ASSERT(cr->nnodes == 1 && cr->sim_nodeid == 0, "Only a single (master) process should execute here");
155 isMasterRankOfNode = true;
158 /* With CUDA detect only on one rank per host, with OpenCL need do
159 * the detection on all PP ranks */
160 bool isOpenclPpRank = ((GMX_GPU == GMX_GPU_OPENCL) && thisRankHasDuty(cr, DUTY_PP));
162 bool gpusCanBeDetected = false;
163 if (isMasterRankOfNode || isOpenclPpRank)
165 gpusCanBeDetected = canDetectGpus();
166 // No need to tell the user anything at this point, they get a
167 // hardware report later.
170 if (gpusCanBeDetected)
172 char detection_error[STRLEN] = "", sbuf[STRLEN];
174 if (detect_gpus(&hwinfo_g->gpu_info, detection_error) != 0)
176 if (detection_error[0] != '\0')
178 sprintf(sbuf, ":\n %s\n", detection_error);
184 GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
185 "NOTE: Error occurred during GPU detection%s"
186 " Can not use GPU acceleration, will fall back to CPU kernels.",
194 /* Broadcast the GPU info to the other ranks within this node */
195 MPI_Bcast(&hwinfo_g->gpu_info.n_dev, 1, MPI_INT, 0, physicalnode_comm);
197 if (hwinfo_g->gpu_info.n_dev > 0)
201 dev_size = hwinfo_g->gpu_info.n_dev*sizeof_gpu_dev_info();
203 if (!isMasterRankOfNode)
205 hwinfo_g->gpu_info.gpu_dev =
206 (struct gmx_device_info_t *)malloc(dev_size);
208 MPI_Bcast(hwinfo_g->gpu_info.gpu_dev, dev_size, MPI_BYTE,
209 0, physicalnode_comm);
210 MPI_Bcast(&hwinfo_g->gpu_info.n_dev_compatible, 1, MPI_INT,
211 0, physicalnode_comm);
215 MPI_Comm_free(&physicalnode_comm);
219 //! Reduce the locally collected \p hwinfo_g over MPI ranks
220 static void gmx_collect_hardware_mpi(const gmx::CpuInfo &cpuInfo)
222 const int ncore = hwinfo_g->hardwareTopology->numberOfCores();
223 /* Zen has family=23, for now we treat future AMD CPUs like Zen */
224 const bool cpuIsAmdZen = (cpuInfo.vendor() == CpuInfo::Vendor::Amd &&
225 cpuInfo.family() >= 23);
229 int nrank, rank, nhwthread, ngpu, i;
233 rank_id = gmx_physicalnode_id_hash();
234 MPI_Comm_rank(MPI_COMM_WORLD, &rank);
235 MPI_Comm_size(MPI_COMM_WORLD, &nrank);
236 nhwthread = hwinfo_g->nthreads_hw_avail;
237 ngpu = hwinfo_g->gpu_info.n_dev_compatible;
238 /* Create a unique hash of the GPU type(s) in this node */
240 /* Here it might be better to only loop over the compatible GPU, but we
241 * don't have that information available and it would also require
242 * removing the device ID from the device info string.
244 for (i = 0; i < hwinfo_g->gpu_info.n_dev; i++)
248 /* Since the device ID is incorporated in the hash, the order of
249 * the GPUs affects the hash. Also two identical GPUs won't give
250 * a gpu_hash of zero after XORing.
252 get_gpu_device_info_string(stmp, hwinfo_g->gpu_info, i);
253 gpu_hash ^= gmx_string_fullhash_func(stmp, gmx_string_hash_init);
260 MPI_Allreduce(buf, all, nrank, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
263 int nnode0, ncore0, nhwthread0, ngpu0, r;
270 for (r = 0; r < nrank; r++)
272 if (all[r] == rank_id)
274 if (!bFound && r == rank)
276 /* We are the first rank in this physical node */
279 nhwthread0 = nhwthread;
289 constexpr int numElementsCounts = 4;
290 std::array<int, numElementsCounts> countsReduced;
292 std::array<int, numElementsCounts> countsLocal;
293 /* Sum values from only intra-rank 0 so we get the sum over all nodes */
294 countsLocal[0] = nnode0;
295 countsLocal[1] = ncore0;
296 countsLocal[2] = nhwthread0;
297 countsLocal[3] = ngpu0;
299 MPI_Allreduce(countsLocal.data(), countsReduced.data(), countsLocal.size(),
300 MPI_INT, MPI_SUM, MPI_COMM_WORLD);
303 constexpr int numElementsMax = 11;
304 std::array<int, numElementsMax> maxMinReduced;
306 std::array<int, numElementsMax> maxMinLocal;
307 /* Store + and - values for all ranks,
308 * so we can get max+min with one MPI call.
310 maxMinLocal[0] = ncore;
311 maxMinLocal[1] = nhwthread;
312 maxMinLocal[2] = ngpu;
313 maxMinLocal[3] = static_cast<int>(gmx::simdSuggested(cpuInfo));
314 maxMinLocal[4] = gpu_hash;
315 maxMinLocal[5] = -maxMinLocal[0];
316 maxMinLocal[6] = -maxMinLocal[1];
317 maxMinLocal[7] = -maxMinLocal[2];
318 maxMinLocal[8] = -maxMinLocal[3];
319 maxMinLocal[9] = -maxMinLocal[4];
320 maxMinLocal[10] = (cpuIsAmdZen ? 1 : 0);
322 MPI_Allreduce(maxMinLocal.data(), maxMinReduced.data(), maxMinLocal.size(),
323 MPI_INT, MPI_MAX, MPI_COMM_WORLD);
326 hwinfo_g->nphysicalnode = countsReduced[0];
327 hwinfo_g->ncore_tot = countsReduced[1];
328 hwinfo_g->ncore_min = -maxMinReduced[5];
329 hwinfo_g->ncore_max = maxMinReduced[0];
330 hwinfo_g->nhwthread_tot = countsReduced[2];
331 hwinfo_g->nhwthread_min = -maxMinReduced[6];
332 hwinfo_g->nhwthread_max = maxMinReduced[1];
333 hwinfo_g->ngpu_compatible_tot = countsReduced[3];
334 hwinfo_g->ngpu_compatible_min = -maxMinReduced[7];
335 hwinfo_g->ngpu_compatible_max = maxMinReduced[2];
336 hwinfo_g->simd_suggest_min = -maxMinReduced[8];
337 hwinfo_g->simd_suggest_max = maxMinReduced[3];
338 hwinfo_g->bIdenticalGPUs = (maxMinReduced[4] == -maxMinReduced[9]);
339 hwinfo_g->haveAmdZenCpu = (maxMinReduced[10] > 0);
341 /* All ranks use the same pointer, protected by a mutex in the caller */
342 hwinfo_g->nphysicalnode = 1;
343 hwinfo_g->ncore_tot = ncore;
344 hwinfo_g->ncore_min = ncore;
345 hwinfo_g->ncore_max = ncore;
346 hwinfo_g->nhwthread_tot = hwinfo_g->nthreads_hw_avail;
347 hwinfo_g->nhwthread_min = hwinfo_g->nthreads_hw_avail;
348 hwinfo_g->nhwthread_max = hwinfo_g->nthreads_hw_avail;
349 hwinfo_g->ngpu_compatible_tot = hwinfo_g->gpu_info.n_dev_compatible;
350 hwinfo_g->ngpu_compatible_min = hwinfo_g->gpu_info.n_dev_compatible;
351 hwinfo_g->ngpu_compatible_max = hwinfo_g->gpu_info.n_dev_compatible;
352 hwinfo_g->simd_suggest_min = static_cast<int>(simdSuggested(cpuInfo));
353 hwinfo_g->simd_suggest_max = static_cast<int>(simdSuggested(cpuInfo));
354 hwinfo_g->bIdenticalGPUs = TRUE;
355 hwinfo_g->haveAmdZenCpu = cpuIsAmdZen;
359 /*! \brief Utility that does dummy computing for max 2 seconds to spin up cores
361 * This routine will check the number of cores configured and online
362 * (using sysconf), and the spins doing dummy compute operations for up to
363 * 2 seconds, or until all cores have come online. This can be used prior to
364 * hardware detection for platforms that take unused processors offline.
366 * This routine will not throw exceptions.
369 spinUpCore() noexcept
371 #if defined(HAVE_SYSCONF) && defined(_SC_NPROCESSORS_CONF) && defined(_SC_NPROCESSORS_ONLN)
373 int countConfigured = sysconf(_SC_NPROCESSORS_CONF); // noexcept
374 auto start = std::chrono::steady_clock::now(); // noexcept
376 while (sysconf(_SC_NPROCESSORS_ONLN) < countConfigured &&
377 std::chrono::steady_clock::now() - start < std::chrono::seconds(2))
379 for (int i = 1; i < 10000; i++)
387 printf("This cannot happen, but prevents loop from being optimized away.");
392 /*! \brief Prepare the system before hardware topology detection
394 * This routine should perform any actions we want to put the system in a state
395 * where we want it to be before detecting the hardware topology. For most
396 * processors there is nothing to do, but some architectures (in particular ARM)
397 * have support for taking configured cores offline, which will make them disappear
398 * from the online processor count.
400 * This routine checks if there is a mismatch between the number of cores
401 * configured and online, and in that case we issue a small workload that
402 * attempts to wake sleeping cores before doing the actual detection.
404 * This type of mismatch can also occur for x86 or PowerPC on Linux, if SMT has only
405 * been disabled in the kernel (rather than bios). Since those cores will never
406 * come online automatically, we currently skip this test for x86 & PowerPC to
407 * avoid wasting 2 seconds. We also skip the test if there is no thread support.
409 * \note Cores will sleep relatively quickly again, so it's important to issue
410 * the real detection code directly after this routine.
413 hardwareTopologyPrepareDetection()
415 #if defined(HAVE_SYSCONF) && defined(_SC_NPROCESSORS_CONF) && \
416 (defined(THREAD_PTHREADS) || defined(THREAD_WINDOWS))
418 // Modify this conditional when/if x86 or PowerPC starts to sleep some cores
419 if (c_architecture != Architecture::X86 &&
420 c_architecture != Architecture::PowerPC)
422 int countConfigured = sysconf(_SC_NPROCESSORS_CONF);
423 std::vector<std::thread> workThreads(countConfigured);
425 for (auto &t : workThreads)
427 t = std::thread(spinUpCore);
430 for (auto &t : workThreads)
438 /*! \brief Sanity check hardware topology and print some notes to log
440 * \param mdlog Logger.
441 * \param hardwareTopology Reference to hardwareTopology object.
444 hardwareTopologyDoubleCheckDetection(const gmx::MDLogger gmx_unused &mdlog,
445 const gmx::HardwareTopology gmx_unused &hardwareTopology)
447 #if defined HAVE_SYSCONF && defined(_SC_NPROCESSORS_CONF)
448 if (hardwareTopology.supportLevel() < gmx::HardwareTopology::SupportLevel::LogicalProcessorCount)
453 int countFromDetection = hardwareTopology.machine().logicalProcessorCount;
454 int countConfigured = sysconf(_SC_NPROCESSORS_CONF);
456 /* BIOS, kernel or user actions can take physical processors
457 * offline. We already cater for the some of the cases inside the hardwareToplogy
458 * by trying to spin up cores just before we detect, but there could be other
459 * cases where it is worthwhile to hint that there might be more resources available.
461 if (countConfigured >= 0 && countConfigured != countFromDetection)
464 appendTextFormatted("Note: %d CPUs configured, but only %d were detected to be online.\n", countConfigured, countFromDetection);
466 if (c_architecture == Architecture::X86 &&
467 countConfigured == 2*countFromDetection)
470 appendText(" X86 Hyperthreading is likely disabled; enable it for better performance.");
472 // For PowerPC (likely Power8) it is possible to set SMT to either 2,4, or 8-way hardware threads.
473 // We only warn if it is completely disabled since default performance drops with SMT8.
474 if (c_architecture == Architecture::PowerPC &&
475 countConfigured == 8*countFromDetection)
478 appendText(" PowerPC SMT is likely disabled; enable SMT2/SMT4 for better performance.");
484 gmx_hw_info_t *gmx_detect_hardware(const gmx::MDLogger &mdlog, const t_commrec *cr)
488 /* make sure no one else is doing the same thing */
489 ret = tMPI_Thread_mutex_lock(&hw_info_lock);
492 gmx_fatal(FARGS, "Error locking hwinfo mutex: %s", strerror(errno));
495 /* only initialize the hwinfo structure if it is not already initalized */
498 hwinfo_g = compat::make_unique<gmx_hw_info_t>();
500 hwinfo_g->cpuInfo = new gmx::CpuInfo(gmx::CpuInfo::detect());
502 hardwareTopologyPrepareDetection();
503 hwinfo_g->hardwareTopology = new gmx::HardwareTopology(gmx::HardwareTopology::detect());
505 // If we detected the topology on this system, double-check that it makes sense
506 if (hwinfo_g->hardwareTopology->isThisSystem())
508 hardwareTopologyDoubleCheckDetection(mdlog, *(hwinfo_g->hardwareTopology));
511 // TODO: Get rid of this altogether.
512 hwinfo_g->nthreads_hw_avail = hwinfo_g->hardwareTopology->machine().logicalProcessorCount;
515 hwinfo_g->gpu_info.n_dev = 0;
516 hwinfo_g->gpu_info.n_dev_compatible = 0;
517 hwinfo_g->gpu_info.gpu_dev = nullptr;
519 gmx_detect_gpus(mdlog, cr);
520 gmx_collect_hardware_mpi(*hwinfo_g->cpuInfo);
522 /* increase the reference counter */
525 ret = tMPI_Thread_mutex_unlock(&hw_info_lock);
528 gmx_fatal(FARGS, "Error unlocking hwinfo mutex: %s", strerror(errno));
531 return hwinfo_g.get();
534 bool compatibleGpusFound(const gmx_gpu_info_t &gpu_info)
536 return gpu_info.n_dev_compatible > 0;
539 void gmx_hardware_info_free()
543 ret = tMPI_Thread_mutex_lock(&hw_info_lock);
546 gmx_fatal(FARGS, "Error locking hwinfo mutex: %s", strerror(errno));
549 /* decrease the reference counter */
555 gmx_incons("n_hwinfo < 0");
560 delete hwinfo_g->cpuInfo;
561 delete hwinfo_g->hardwareTopology;
562 free_gpu_info(&hwinfo_g->gpu_info);
566 ret = tMPI_Thread_mutex_unlock(&hw_info_lock);
569 gmx_fatal(FARGS, "Error unlocking hwinfo mutex: %s", strerror(errno));