2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015,2016 by the GROMACS development team.
5 * Copyright (c) 2017,2018,2019,2020, by the GROMACS development team, led by
6 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
7 * and including many others, as listed in the AUTHORS file in the
8 * top-level source directory and at http://www.gromacs.org.
10 * GROMACS is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public License
12 * as published by the Free Software Foundation; either version 2.1
13 * of the License, or (at your option) any later version.
15 * GROMACS is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with GROMACS; if not, see
22 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
23 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
25 * If you want to redistribute modifications to GROMACS, please
26 * consider that scientific software is very special. Version
27 * control is crucial - bugs must be traceable. We will be happy to
28 * consider code for inclusion in the official distribution, but
29 * derived work must not be called official GROMACS. Details are found
30 * in the README & COPYING files - if they are missing, get the
31 * official version at http://www.gromacs.org.
33 * To help us fund GROMACS development, we humbly ask that you cite
34 * the research papers on the package. Check out http://www.gromacs.org.
38 #include "detecthardware.h"
50 #include "gromacs/compat/pointers.h"
51 #include "gromacs/hardware/cpuinfo.h"
52 #include "gromacs/hardware/device_management.h"
53 #include "gromacs/hardware/hardwaretopology.h"
54 #include "gromacs/hardware/hw_info.h"
55 #include "gromacs/simd/support.h"
56 #include "gromacs/utility/basedefinitions.h"
57 #include "gromacs/utility/basenetwork.h"
58 #include "gromacs/utility/baseversion.h"
59 #include "gromacs/utility/cstringutil.h"
60 #include "gromacs/utility/exceptions.h"
61 #include "gromacs/utility/fatalerror.h"
62 #include "gromacs/utility/gmxassert.h"
63 #include "gromacs/utility/inmemoryserializer.h"
64 #include "gromacs/utility/logger.h"
65 #include "gromacs/utility/mutex.h"
66 #include "gromacs/utility/physicalnodecommunicator.h"
68 #include "architecture.h"
69 #include "device_information.h"
72 # include <unistd.h> // sysconf()
75 gmx_hw_info_t::gmx_hw_info_t(std::unique_ptr<gmx::CpuInfo> cpuInfo,
76 std::unique_ptr<gmx::HardwareTopology> hardwareTopology) :
77 cpuInfo(std::move(cpuInfo)),
78 hardwareTopology(std::move(hardwareTopology))
82 gmx_hw_info_t::~gmx_hw_info_t() = default;
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 std::string errorMessage;
115 if (!canPerformDeviceDetection(&errorMessage))
119 .appendTextFormatted(
120 "NOTE: Detection of GPUs failed. The API reported:\n"
122 " GROMACS cannot run tasks on a GPU.",
123 errorMessage.c_str());
127 bool isMasterRankOfPhysicalNode = true;
129 isMasterRankOfPhysicalNode = (physicalNodeComm.rank_ == 0);
131 // We choose to run the detection only once with thread-MPI and
132 // use a mutex to enforce it.
133 GMX_UNUSED_VALUE(physicalNodeComm);
134 isMasterRankOfPhysicalNode = true;
137 /* The OpenCL support requires us to run detection on all ranks.
138 * With CUDA we don't need to, and prefer to detect on one rank
139 * and send the information to the other ranks over MPI. */
140 constexpr bool allRanksMustDetectGpus = (GMX_GPU_OPENCL != 0);
141 bool gpusCanBeDetected = false;
142 if (isMasterRankOfPhysicalNode || allRanksMustDetectGpus)
144 std::string errorMessage;
145 gpusCanBeDetected = isDeviceDetectionFunctional(&errorMessage);
146 if (!gpusCanBeDetected)
150 .appendTextFormatted(
151 "NOTE: Detection of GPUs failed. The API reported:\n"
153 " GROMACS cannot run tasks on a GPU.",
154 errorMessage.c_str());
158 if (gpusCanBeDetected)
160 hardwareInfo->deviceInfoList = findDevices();
161 // No need to tell the user anything at this point, they get a
162 // hardware report later.
166 if (!allRanksMustDetectGpus && !hardwareInfo->deviceInfoList.empty())
168 gmx::InMemorySerializer writer;
169 serializeDeviceInformations(hardwareInfo->deviceInfoList, &writer);
170 auto buffer = writer.finishAndGetBuffer();
172 MPI_Bcast(buffer.data(), buffer.size(), MPI_BYTE, 0, physicalNodeComm.comm_);
174 gmx::InMemoryDeserializer reader(buffer, false);
175 hardwareInfo->deviceInfoList = deserializeDeviceInformations(&writer);
180 //! Reduce the locally collected \p hardwareInfo over MPI ranks
181 static void gmx_collect_hardware_mpi(const gmx::CpuInfo& cpuInfo,
182 const PhysicalNodeCommunicator& physicalNodeComm,
183 compat::not_null<gmx_hw_info_t*> hardwareInfo)
185 const int ncore = hardwareInfo->hardwareTopology->numberOfCores();
186 /* Zen1 is assumed for:
187 * - family=23 with the below listed models;
190 const bool cpuIsAmdZen1 = ((cpuInfo.vendor() == CpuInfo::Vendor::Amd && cpuInfo.family() == 23
191 && (cpuInfo.model() == 1 || cpuInfo.model() == 17
192 || cpuInfo.model() == 8 || cpuInfo.model() == 24))
193 || cpuInfo.vendor() == CpuInfo::Vendor::Hygon);
195 int numCompatibleDevices = getCompatibleDevices(hardwareInfo->deviceInfoList).size();
200 nhwthread = hardwareInfo->nthreads_hw_avail;
201 /* Create a unique hash of the GPU type(s) in this node */
203 /* Here it might be better to only loop over the compatible GPU, but we
204 * don't have that information available and it would also require
205 * removing the device ID from the device info string.
207 for (const auto& deviceInfo : hardwareInfo->deviceInfoList)
209 /* Since the device ID is incorporated in the hash, the order of
210 * the GPUs affects the hash. Also two identical GPUs won't give
211 * a gpu_hash of zero after XORing.
213 std::string deviceInfoString = getDeviceInformationString(*deviceInfo);
214 gpu_hash ^= gmx_string_fullhash_func(deviceInfoString.c_str(), gmx_string_hash_init);
217 constexpr int numElementsCounts = 4;
218 std::array<int, numElementsCounts> countsReduced;
220 std::array<int, numElementsCounts> countsLocal = { { 0 } };
221 // Organize to sum values from only one rank within each node,
222 // so we get the sum over all nodes.
223 bool isMasterRankOfPhysicalNode = (physicalNodeComm.rank_ == 0);
224 if (isMasterRankOfPhysicalNode)
227 countsLocal[1] = ncore;
228 countsLocal[2] = nhwthread;
229 countsLocal[3] = numCompatibleDevices;
232 MPI_Allreduce(countsLocal.data(), countsReduced.data(), countsLocal.size(), MPI_INT,
233 MPI_SUM, MPI_COMM_WORLD);
236 constexpr int numElementsMax = 11;
237 std::array<int, numElementsMax> maxMinReduced;
239 std::array<int, numElementsMax> maxMinLocal;
240 /* Store + and - values for all ranks,
241 * so we can get max+min with one MPI call.
243 maxMinLocal[0] = ncore;
244 maxMinLocal[1] = nhwthread;
245 maxMinLocal[2] = numCompatibleDevices;
246 maxMinLocal[3] = static_cast<int>(gmx::simdSuggested(cpuInfo));
247 maxMinLocal[4] = gpu_hash;
248 maxMinLocal[5] = -maxMinLocal[0];
249 maxMinLocal[6] = -maxMinLocal[1];
250 maxMinLocal[7] = -maxMinLocal[2];
251 maxMinLocal[8] = -maxMinLocal[3];
252 maxMinLocal[9] = -maxMinLocal[4];
253 maxMinLocal[10] = (cpuIsAmdZen1 ? 1 : 0);
255 MPI_Allreduce(maxMinLocal.data(), maxMinReduced.data(), maxMinLocal.size(), MPI_INT,
256 MPI_MAX, MPI_COMM_WORLD);
259 hardwareInfo->nphysicalnode = countsReduced[0];
260 hardwareInfo->ncore_tot = countsReduced[1];
261 hardwareInfo->ncore_min = -maxMinReduced[5];
262 hardwareInfo->ncore_max = maxMinReduced[0];
263 hardwareInfo->nhwthread_tot = countsReduced[2];
264 hardwareInfo->nhwthread_min = -maxMinReduced[6];
265 hardwareInfo->nhwthread_max = maxMinReduced[1];
266 hardwareInfo->ngpu_compatible_tot = countsReduced[3];
267 hardwareInfo->ngpu_compatible_min = -maxMinReduced[7];
268 hardwareInfo->ngpu_compatible_max = maxMinReduced[2];
269 hardwareInfo->simd_suggest_min = -maxMinReduced[8];
270 hardwareInfo->simd_suggest_max = maxMinReduced[3];
271 hardwareInfo->bIdenticalGPUs = (maxMinReduced[4] == -maxMinReduced[9]);
272 hardwareInfo->haveAmdZen1Cpu = (maxMinReduced[10] > 0);
274 /* All ranks use the same pointer, protected by a mutex in the caller */
275 hardwareInfo->nphysicalnode = 1;
276 hardwareInfo->ncore_tot = ncore;
277 hardwareInfo->ncore_min = ncore;
278 hardwareInfo->ncore_max = ncore;
279 hardwareInfo->nhwthread_tot = hardwareInfo->nthreads_hw_avail;
280 hardwareInfo->nhwthread_min = hardwareInfo->nthreads_hw_avail;
281 hardwareInfo->nhwthread_max = hardwareInfo->nthreads_hw_avail;
282 hardwareInfo->ngpu_compatible_tot = numCompatibleDevices;
283 hardwareInfo->ngpu_compatible_min = numCompatibleDevices;
284 hardwareInfo->ngpu_compatible_max = numCompatibleDevices;
285 hardwareInfo->simd_suggest_min = static_cast<int>(simdSuggested(cpuInfo));
286 hardwareInfo->simd_suggest_max = static_cast<int>(simdSuggested(cpuInfo));
287 hardwareInfo->bIdenticalGPUs = TRUE;
288 hardwareInfo->haveAmdZen1Cpu = cpuIsAmdZen1;
289 GMX_UNUSED_VALUE(physicalNodeComm);
293 /*! \brief Utility that does dummy computing for max 2 seconds to spin up cores
295 * This routine will check the number of cores configured and online
296 * (using sysconf), and the spins doing dummy compute operations for up to
297 * 2 seconds, or until all cores have come online. This can be used prior to
298 * hardware detection for platforms that take unused processors offline.
300 * This routine will not throw exceptions. In principle it should be
301 * declared noexcept, but at least icc 19.1 and 21-beta08 with the
302 * libstdc++-7.5 has difficulty implementing a std::vector of
303 * std::thread started with this function when declared noexcept. It
304 * is not clear whether the problem is the compiler or the standard
305 * library. Fortunately, this function is not performance sensitive,
306 * and only runs on platforms other than x86 and POWER (ie ARM),
307 * so the possible overhead introduced by omitting noexcept is not
310 static void spinUpCore()
312 #if defined(HAVE_SYSCONF) && defined(_SC_NPROCESSORS_CONF) && defined(_SC_NPROCESSORS_ONLN)
314 int countConfigured = sysconf(_SC_NPROCESSORS_CONF); // noexcept
315 auto start = std::chrono::steady_clock::now(); // noexcept
317 while (sysconf(_SC_NPROCESSORS_ONLN) < countConfigured
318 && std::chrono::steady_clock::now() - start < std::chrono::seconds(2))
320 for (int i = 1; i < 10000; i++)
328 printf("This cannot happen, but prevents loop from being optimized away.");
333 /*! \brief Prepare the system before hardware topology detection
335 * This routine should perform any actions we want to put the system in a state
336 * where we want it to be before detecting the hardware topology. For most
337 * processors there is nothing to do, but some architectures (in particular ARM)
338 * have support for taking configured cores offline, which will make them disappear
339 * from the online processor count.
341 * This routine checks if there is a mismatch between the number of cores
342 * configured and online, and in that case we issue a small workload that
343 * attempts to wake sleeping cores before doing the actual detection.
345 * This type of mismatch can also occur for x86 or PowerPC on Linux, if SMT has only
346 * been disabled in the kernel (rather than bios). Since those cores will never
347 * come online automatically, we currently skip this test for x86 & PowerPC to
348 * avoid wasting 2 seconds. We also skip the test if there is no thread support.
350 * \note Cores will sleep relatively quickly again, so it's important to issue
351 * the real detection code directly after this routine.
353 static void hardwareTopologyPrepareDetection()
355 #if defined(HAVE_SYSCONF) && defined(_SC_NPROCESSORS_CONF) \
356 && (defined(THREAD_PTHREADS) || defined(THREAD_WINDOWS))
358 // Modify this conditional when/if x86 or PowerPC starts to sleep some cores
359 if (c_architecture != Architecture::X86 && c_architecture != Architecture::PowerPC)
361 int countConfigured = sysconf(_SC_NPROCESSORS_CONF);
362 std::vector<std::thread> workThreads(countConfigured);
364 for (auto& t : workThreads)
366 t = std::thread(spinUpCore);
369 for (auto& t : workThreads)
377 /*! \brief Sanity check hardware topology and print some notes to log
379 * \param mdlog Logger.
380 * \param hardwareTopology Reference to hardwareTopology object.
382 static void hardwareTopologyDoubleCheckDetection(const gmx::MDLogger gmx_unused& mdlog,
383 const gmx::HardwareTopology gmx_unused& hardwareTopology)
385 #if defined HAVE_SYSCONF && defined(_SC_NPROCESSORS_CONF)
386 if (hardwareTopology.supportLevel() < gmx::HardwareTopology::SupportLevel::LogicalProcessorCount)
391 int countFromDetection = hardwareTopology.machine().logicalProcessorCount;
392 int countConfigured = sysconf(_SC_NPROCESSORS_CONF);
394 /* BIOS, kernel or user actions can take physical processors
395 * offline. We already cater for the some of the cases inside the hardwareToplogy
396 * by trying to spin up cores just before we detect, but there could be other
397 * cases where it is worthwhile to hint that there might be more resources available.
399 if (countConfigured >= 0 && countConfigured != countFromDetection)
402 .appendTextFormatted(
403 "Note: %d CPUs configured, but only %d were detected to be online.\n",
404 countConfigured, countFromDetection);
406 if (c_architecture == Architecture::X86 && countConfigured == 2 * countFromDetection)
410 " X86 Hyperthreading is likely disabled; enable it for better "
413 // For PowerPC (likely Power8) it is possible to set SMT to either 2,4, or 8-way hardware threads.
414 // We only warn if it is completely disabled since default performance drops with SMT8.
415 if (c_architecture == Architecture::PowerPC && countConfigured == 8 * countFromDetection)
419 " PowerPC SMT is likely disabled; enable SMT2/SMT4 for better "
426 gmx_hw_info_t* gmx_detect_hardware(const gmx::MDLogger& mdlog, const PhysicalNodeCommunicator& physicalNodeComm)
428 // By construction, only one thread ever runs hardware detection,
429 // but we may as well prevent issues arising if that would change.
430 // Taking the lock early ensures that exactly one thread can
431 // attempt to construct g_hardwareInfo.
432 lock_guard<Mutex> lock(g_hardwareInfoMutex);
434 // If we already have the information, just return a handle to it.
435 if (g_hardwareInfo != nullptr)
437 return g_hardwareInfo.get();
440 // Make the new hardwareInfo in a temporary.
441 hardwareTopologyPrepareDetection();
443 // TODO: We should also do CPU hardware detection only once on each
444 // physical node and broadcast it, instead of doing it on every MPI rank.
445 auto hardwareInfo = std::make_unique<gmx_hw_info_t>(
446 std::make_unique<CpuInfo>(CpuInfo::detect()),
447 std::make_unique<HardwareTopology>(HardwareTopology::detect()));
449 // If we detected the topology on this system, double-check that it makes sense
450 if (hardwareInfo->hardwareTopology->isThisSystem())
452 hardwareTopologyDoubleCheckDetection(mdlog, *hardwareInfo->hardwareTopology);
455 // TODO: Get rid of this altogether.
456 hardwareInfo->nthreads_hw_avail = hardwareInfo->hardwareTopology->machine().logicalProcessorCount;
459 gmx_detect_gpus(mdlog, physicalNodeComm, compat::make_not_null(hardwareInfo));
460 gmx_collect_hardware_mpi(*hardwareInfo->cpuInfo, physicalNodeComm, compat::make_not_null(hardwareInfo));
462 // Now that the temporary is fully constructed, swap it to become
464 g_hardwareInfo.swap(hardwareInfo);
466 return g_hardwareInfo.get();