2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015,2016,2017,2018,2019, 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"
49 #include "gromacs/compat/pointers.h"
50 #include "gromacs/gpu_utils/gpu_utils.h"
51 #include "gromacs/hardware/cpuinfo.h"
52 #include "gromacs/hardware/hardwaretopology.h"
53 #include "gromacs/hardware/hw_info.h"
54 #include "gromacs/simd/support.h"
55 #include "gromacs/utility/basedefinitions.h"
56 #include "gromacs/utility/basenetwork.h"
57 #include "gromacs/utility/baseversion.h"
58 #include "gromacs/utility/cstringutil.h"
59 #include "gromacs/utility/exceptions.h"
60 #include "gromacs/utility/fatalerror.h"
61 #include "gromacs/utility/gmxassert.h"
62 #include "gromacs/utility/logger.h"
63 #include "gromacs/utility/mutex.h"
64 #include "gromacs/utility/physicalnodecommunicator.h"
66 #include "architecture.h"
69 # include <unistd.h> // sysconf()
72 gmx_hw_info_t::gmx_hw_info_t(std::unique_ptr<gmx::CpuInfo> cpuInfo,
73 std::unique_ptr<gmx::HardwareTopology> hardwareTopology) :
74 cpuInfo(std::move(cpuInfo)),
75 hardwareTopology(std::move(hardwareTopology))
79 gmx_hw_info_t::~gmx_hw_info_t()
81 free_gpu_info(&gpu_info);
87 //! Convenience macro to help us avoid ifdefs each time we use sysconf
88 #if !defined(_SC_NPROCESSORS_ONLN) && defined(_SC_NPROC_ONLN)
89 # define _SC_NPROCESSORS_ONLN _SC_NPROC_ONLN
92 //! Convenience macro to help us avoid ifdefs each time we use sysconf
93 #if !defined(_SC_NPROCESSORS_CONF) && defined(_SC_NPROC_CONF)
94 # define _SC_NPROCESSORS_CONF _SC_NPROC_CONF
97 /*! \brief Information about the hardware of all nodes (common to all threads in this process).
99 * This information is constructed only when required, but thereafter
100 * its lifetime is that of the whole process, potentially across
101 * multiple successive simulation parts. It's wise to ensure that only
102 * one thread can create the information, but thereafter they can all
103 * read it without e.g. needing a std::shared_ptr to ensure its
104 * lifetime exceeds that of the thread. */
105 static std::unique_ptr<gmx_hw_info_t> g_hardwareInfo;
106 //! A mutex to protect the hwinfo structure
107 static Mutex g_hardwareInfoMutex;
109 //! Detect GPUs, if that makes sense to attempt.
110 static void gmx_detect_gpus(const gmx::MDLogger& mdlog,
111 const PhysicalNodeCommunicator& physicalNodeComm,
112 compat::not_null<gmx_hw_info_t*> hardwareInfo)
114 hardwareInfo->gpu_info.bDetectGPUs = canPerformGpuDetection();
116 if (!hardwareInfo->gpu_info.bDetectGPUs)
121 bool isMasterRankOfPhysicalNode = true;
123 isMasterRankOfPhysicalNode = (physicalNodeComm.rank_ == 0);
125 // We choose to run the detection only once with thread-MPI and
126 // use a mutex to enforce it.
127 GMX_UNUSED_VALUE(physicalNodeComm);
128 isMasterRankOfPhysicalNode = true;
131 /* The OpenCL support requires us to run detection on all ranks.
132 * With CUDA we don't need to, and prefer to detect on one rank
133 * and send the information to the other ranks over MPI. */
134 bool allRanksMustDetectGpus = (GMX_GPU == GMX_GPU_OPENCL);
135 bool gpusCanBeDetected = false;
136 if (isMasterRankOfPhysicalNode || allRanksMustDetectGpus)
138 std::string errorMessage;
139 gpusCanBeDetected = isGpuDetectionFunctional(&errorMessage);
140 if (!gpusCanBeDetected)
144 .appendTextFormatted(
145 "NOTE: Detection of GPUs failed. The API reported:\n"
147 " GROMACS cannot run tasks on a GPU.",
148 errorMessage.c_str());
152 if (gpusCanBeDetected)
154 findGpus(&hardwareInfo->gpu_info);
155 // No need to tell the user anything at this point, they get a
156 // hardware report later.
160 if (!allRanksMustDetectGpus)
162 /* Broadcast the GPU info to the other ranks within this node */
163 MPI_Bcast(&hardwareInfo->gpu_info.n_dev, 1, MPI_INT, 0, physicalNodeComm.comm_);
165 if (hardwareInfo->gpu_info.n_dev > 0)
169 dev_size = hardwareInfo->gpu_info.n_dev * sizeof_gpu_dev_info();
171 if (!isMasterRankOfPhysicalNode)
173 hardwareInfo->gpu_info.gpu_dev = (struct gmx_device_info_t*)malloc(dev_size);
175 MPI_Bcast(hardwareInfo->gpu_info.gpu_dev, dev_size, MPI_BYTE, 0, physicalNodeComm.comm_);
176 MPI_Bcast(&hardwareInfo->gpu_info.n_dev_compatible, 1, MPI_INT, 0, physicalNodeComm.comm_);
182 //! Reduce the locally collected \p hardwareInfo over MPI ranks
183 static void gmx_collect_hardware_mpi(const gmx::CpuInfo& cpuInfo,
184 const PhysicalNodeCommunicator& physicalNodeComm,
185 compat::not_null<gmx_hw_info_t*> hardwareInfo)
187 const int ncore = hardwareInfo->hardwareTopology->numberOfCores();
188 /* Zen1 is assumed for:
189 * - family=23 with the below listed models;
192 const bool cpuIsAmdZen1 = ((cpuInfo.vendor() == CpuInfo::Vendor::Amd && cpuInfo.family() == 23
193 && (cpuInfo.model() == 1 || cpuInfo.model() == 17
194 || cpuInfo.model() == 8 || cpuInfo.model() == 24))
195 || cpuInfo.vendor() == CpuInfo::Vendor::Hygon);
197 int nhwthread, ngpu, i;
200 nhwthread = hardwareInfo->nthreads_hw_avail;
201 ngpu = hardwareInfo->gpu_info.n_dev_compatible;
202 /* Create a unique hash of the GPU type(s) in this node */
204 /* Here it might be better to only loop over the compatible GPU, but we
205 * don't have that information available and it would also require
206 * removing the device ID from the device info string.
208 for (i = 0; i < hardwareInfo->gpu_info.n_dev; i++)
212 /* Since the device ID is incorporated in the hash, the order of
213 * the GPUs affects the hash. Also two identical GPUs won't give
214 * a gpu_hash of zero after XORing.
216 get_gpu_device_info_string(stmp, hardwareInfo->gpu_info, i);
217 gpu_hash ^= gmx_string_fullhash_func(stmp, gmx_string_hash_init);
220 constexpr int numElementsCounts = 4;
221 std::array<int, numElementsCounts> countsReduced;
223 std::array<int, numElementsCounts> countsLocal = { { 0 } };
224 // Organize to sum values from only one rank within each node,
225 // so we get the sum over all nodes.
226 bool isMasterRankOfPhysicalNode = (physicalNodeComm.rank_ == 0);
227 if (isMasterRankOfPhysicalNode)
230 countsLocal[1] = ncore;
231 countsLocal[2] = nhwthread;
232 countsLocal[3] = ngpu;
235 MPI_Allreduce(countsLocal.data(), countsReduced.data(), countsLocal.size(), MPI_INT,
236 MPI_SUM, MPI_COMM_WORLD);
239 constexpr int numElementsMax = 11;
240 std::array<int, numElementsMax> maxMinReduced;
242 std::array<int, numElementsMax> maxMinLocal;
243 /* Store + and - values for all ranks,
244 * so we can get max+min with one MPI call.
246 maxMinLocal[0] = ncore;
247 maxMinLocal[1] = nhwthread;
248 maxMinLocal[2] = ngpu;
249 maxMinLocal[3] = static_cast<int>(gmx::simdSuggested(cpuInfo));
250 maxMinLocal[4] = gpu_hash;
251 maxMinLocal[5] = -maxMinLocal[0];
252 maxMinLocal[6] = -maxMinLocal[1];
253 maxMinLocal[7] = -maxMinLocal[2];
254 maxMinLocal[8] = -maxMinLocal[3];
255 maxMinLocal[9] = -maxMinLocal[4];
256 maxMinLocal[10] = (cpuIsAmdZen1 ? 1 : 0);
258 MPI_Allreduce(maxMinLocal.data(), maxMinReduced.data(), maxMinLocal.size(), MPI_INT,
259 MPI_MAX, MPI_COMM_WORLD);
262 hardwareInfo->nphysicalnode = countsReduced[0];
263 hardwareInfo->ncore_tot = countsReduced[1];
264 hardwareInfo->ncore_min = -maxMinReduced[5];
265 hardwareInfo->ncore_max = maxMinReduced[0];
266 hardwareInfo->nhwthread_tot = countsReduced[2];
267 hardwareInfo->nhwthread_min = -maxMinReduced[6];
268 hardwareInfo->nhwthread_max = maxMinReduced[1];
269 hardwareInfo->ngpu_compatible_tot = countsReduced[3];
270 hardwareInfo->ngpu_compatible_min = -maxMinReduced[7];
271 hardwareInfo->ngpu_compatible_max = maxMinReduced[2];
272 hardwareInfo->simd_suggest_min = -maxMinReduced[8];
273 hardwareInfo->simd_suggest_max = maxMinReduced[3];
274 hardwareInfo->bIdenticalGPUs = (maxMinReduced[4] == -maxMinReduced[9]);
275 hardwareInfo->haveAmdZen1Cpu = (maxMinReduced[10] > 0);
277 /* All ranks use the same pointer, protected by a mutex in the caller */
278 hardwareInfo->nphysicalnode = 1;
279 hardwareInfo->ncore_tot = ncore;
280 hardwareInfo->ncore_min = ncore;
281 hardwareInfo->ncore_max = ncore;
282 hardwareInfo->nhwthread_tot = hardwareInfo->nthreads_hw_avail;
283 hardwareInfo->nhwthread_min = hardwareInfo->nthreads_hw_avail;
284 hardwareInfo->nhwthread_max = hardwareInfo->nthreads_hw_avail;
285 hardwareInfo->ngpu_compatible_tot = hardwareInfo->gpu_info.n_dev_compatible;
286 hardwareInfo->ngpu_compatible_min = hardwareInfo->gpu_info.n_dev_compatible;
287 hardwareInfo->ngpu_compatible_max = hardwareInfo->gpu_info.n_dev_compatible;
288 hardwareInfo->simd_suggest_min = static_cast<int>(simdSuggested(cpuInfo));
289 hardwareInfo->simd_suggest_max = static_cast<int>(simdSuggested(cpuInfo));
290 hardwareInfo->bIdenticalGPUs = TRUE;
291 hardwareInfo->haveAmdZen1Cpu = cpuIsAmdZen1;
292 GMX_UNUSED_VALUE(physicalNodeComm);
296 /*! \brief Utility that does dummy computing for max 2 seconds to spin up cores
298 * This routine will check the number of cores configured and online
299 * (using sysconf), and the spins doing dummy compute operations for up to
300 * 2 seconds, or until all cores have come online. This can be used prior to
301 * hardware detection for platforms that take unused processors offline.
303 * This routine will not throw exceptions.
305 static void spinUpCore() noexcept
307 #if defined(HAVE_SYSCONF) && defined(_SC_NPROCESSORS_CONF) && defined(_SC_NPROCESSORS_ONLN)
309 int countConfigured = sysconf(_SC_NPROCESSORS_CONF); // noexcept
310 auto start = std::chrono::steady_clock::now(); // noexcept
312 while (sysconf(_SC_NPROCESSORS_ONLN) < countConfigured
313 && std::chrono::steady_clock::now() - start < std::chrono::seconds(2))
315 for (int i = 1; i < 10000; i++)
323 printf("This cannot happen, but prevents loop from being optimized away.");
328 /*! \brief Prepare the system before hardware topology detection
330 * This routine should perform any actions we want to put the system in a state
331 * where we want it to be before detecting the hardware topology. For most
332 * processors there is nothing to do, but some architectures (in particular ARM)
333 * have support for taking configured cores offline, which will make them disappear
334 * from the online processor count.
336 * This routine checks if there is a mismatch between the number of cores
337 * configured and online, and in that case we issue a small workload that
338 * attempts to wake sleeping cores before doing the actual detection.
340 * This type of mismatch can also occur for x86 or PowerPC on Linux, if SMT has only
341 * been disabled in the kernel (rather than bios). Since those cores will never
342 * come online automatically, we currently skip this test for x86 & PowerPC to
343 * avoid wasting 2 seconds. We also skip the test if there is no thread support.
345 * \note Cores will sleep relatively quickly again, so it's important to issue
346 * the real detection code directly after this routine.
348 static void hardwareTopologyPrepareDetection()
350 #if defined(HAVE_SYSCONF) && defined(_SC_NPROCESSORS_CONF) \
351 && (defined(THREAD_PTHREADS) || defined(THREAD_WINDOWS))
353 // Modify this conditional when/if x86 or PowerPC starts to sleep some cores
354 if (c_architecture != Architecture::X86 && c_architecture != Architecture::PowerPC)
356 int countConfigured = sysconf(_SC_NPROCESSORS_CONF);
357 std::vector<std::thread> workThreads(countConfigured);
359 for (auto& t : workThreads)
361 t = std::thread(spinUpCore);
364 for (auto& t : workThreads)
372 /*! \brief Sanity check hardware topology and print some notes to log
374 * \param mdlog Logger.
375 * \param hardwareTopology Reference to hardwareTopology object.
377 static void hardwareTopologyDoubleCheckDetection(const gmx::MDLogger gmx_unused& mdlog,
378 const gmx::HardwareTopology gmx_unused& hardwareTopology)
380 #if defined HAVE_SYSCONF && defined(_SC_NPROCESSORS_CONF)
381 if (hardwareTopology.supportLevel() < gmx::HardwareTopology::SupportLevel::LogicalProcessorCount)
386 int countFromDetection = hardwareTopology.machine().logicalProcessorCount;
387 int countConfigured = sysconf(_SC_NPROCESSORS_CONF);
389 /* BIOS, kernel or user actions can take physical processors
390 * offline. We already cater for the some of the cases inside the hardwareToplogy
391 * by trying to spin up cores just before we detect, but there could be other
392 * cases where it is worthwhile to hint that there might be more resources available.
394 if (countConfigured >= 0 && countConfigured != countFromDetection)
397 .appendTextFormatted(
398 "Note: %d CPUs configured, but only %d were detected to be online.\n",
399 countConfigured, countFromDetection);
401 if (c_architecture == Architecture::X86 && countConfigured == 2 * countFromDetection)
405 " X86 Hyperthreading is likely disabled; enable it for better "
408 // For PowerPC (likely Power8) it is possible to set SMT to either 2,4, or 8-way hardware threads.
409 // We only warn if it is completely disabled since default performance drops with SMT8.
410 if (c_architecture == Architecture::PowerPC && countConfigured == 8 * countFromDetection)
414 " PowerPC SMT is likely disabled; enable SMT2/SMT4 for better "
421 gmx_hw_info_t* gmx_detect_hardware(const gmx::MDLogger& mdlog, const PhysicalNodeCommunicator& physicalNodeComm)
423 // By construction, only one thread ever runs hardware detection,
424 // but we may as well prevent issues arising if that would change.
425 // Taking the lock early ensures that exactly one thread can
426 // attempt to construct g_hardwareInfo.
427 lock_guard<Mutex> lock(g_hardwareInfoMutex);
429 // If we already have the information, just return a handle to it.
430 if (g_hardwareInfo != nullptr)
432 return g_hardwareInfo.get();
435 // Make the new hardwareInfo in a temporary.
436 hardwareTopologyPrepareDetection();
438 // TODO: We should also do CPU hardware detection only once on each
439 // physical node and broadcast it, instead of doing it on every MPI rank.
440 auto hardwareInfo = std::make_unique<gmx_hw_info_t>(
441 std::make_unique<CpuInfo>(CpuInfo::detect()),
442 std::make_unique<HardwareTopology>(HardwareTopology::detect()));
444 // If we detected the topology on this system, double-check that it makes sense
445 if (hardwareInfo->hardwareTopology->isThisSystem())
447 hardwareTopologyDoubleCheckDetection(mdlog, *hardwareInfo->hardwareTopology);
450 // TODO: Get rid of this altogether.
451 hardwareInfo->nthreads_hw_avail = hardwareInfo->hardwareTopology->machine().logicalProcessorCount;
454 hardwareInfo->gpu_info.n_dev = 0;
455 hardwareInfo->gpu_info.n_dev_compatible = 0;
456 hardwareInfo->gpu_info.gpu_dev = nullptr;
458 gmx_detect_gpus(mdlog, physicalNodeComm, compat::make_not_null(hardwareInfo));
459 gmx_collect_hardware_mpi(*hardwareInfo->cpuInfo, physicalNodeComm, compat::make_not_null(hardwareInfo));
461 // Now that the temporary is fully constructed, swap it to become
463 g_hardwareInfo.swap(hardwareInfo);
465 return g_hardwareInfo.get();
468 bool compatibleGpusFound(const gmx_gpu_info_t& gpu_info)
470 return gpu_info.n_dev_compatible > 0;