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"
52 #include "thread_mpi/threads.h"
54 #include "gromacs/compat/make_unique.h"
55 #include "gromacs/gpu_utils/gpu_utils.h"
56 #include "gromacs/hardware/cpuinfo.h"
57 #include "gromacs/hardware/hardwaretopology.h"
58 #include "gromacs/hardware/hw_info.h"
59 #include "gromacs/mdtypes/commrec.h"
60 #include "gromacs/simd/support.h"
61 #include "gromacs/utility/basedefinitions.h"
62 #include "gromacs/utility/basenetwork.h"
63 #include "gromacs/utility/baseversion.h"
64 #include "gromacs/utility/cstringutil.h"
65 #include "gromacs/utility/exceptions.h"
66 #include "gromacs/utility/fatalerror.h"
67 #include "gromacs/utility/gmxassert.h"
68 #include "gromacs/utility/logger.h"
69 #include "gromacs/utility/programcontext.h"
70 #include "gromacs/utility/smalloc.h"
71 #include "gromacs/utility/stringutil.h"
72 #include "gromacs/utility/sysinfo.h"
74 #include "architecture.h"
77 # include <unistd.h> // sysconf()
83 //! Convenience macro to help us avoid ifdefs each time we use sysconf
84 #if !defined(_SC_NPROCESSORS_ONLN) && defined(_SC_NPROC_ONLN)
85 # define _SC_NPROCESSORS_ONLN _SC_NPROC_ONLN
88 //! Convenience macro to help us avoid ifdefs each time we use sysconf
89 #if !defined(_SC_NPROCESSORS_CONF) && defined(_SC_NPROC_CONF)
90 # define _SC_NPROCESSORS_CONF _SC_NPROC_CONF
93 //! Constant used to help minimize preprocessed code
94 static const bool bGPUBinary = GMX_GPU != GMX_GPU_NONE;
96 /*! \brief The hwinfo structure (common to all threads in this process).
98 * \todo This should become a shared_ptr owned by e.g. Mdrunner::runner()
99 * that is shared across any threads as needed (e.g. for thread-MPI). That
100 * offers about the same run time performance as we get here, and avoids a
101 * lot of custom code.
103 static std::unique_ptr<gmx_hw_info_t> hwinfo_g;
104 //! A reference counter for the hwinfo structure
105 static int n_hwinfo = 0;
106 //! A lock to protect the hwinfo structure
107 static tMPI_Thread_mutex_t hw_info_lock = TMPI_THREAD_MUTEX_INITIALIZER;
109 //! Detect GPUs, if that makes sense to attempt.
110 static void gmx_detect_gpus(const gmx::MDLogger &mdlog, const t_commrec *cr)
114 MPI_Comm physicalnode_comm;
118 hwinfo_g->gpu_info.bDetectGPUs =
119 (bGPUBinary && getenv("GMX_DISABLE_GPU_DETECTION") == nullptr);
120 if (!hwinfo_g->gpu_info.bDetectGPUs)
125 /* Under certain circumstances MPI ranks on the same physical node
126 * can not simultaneously access the same GPU(s). Therefore we run
127 * the detection only on one MPI rank per node and broadcast the info.
128 * Note that with thread-MPI only a single thread runs this code.
130 * NOTE: We can't broadcast gpu_info with OpenCL as the device and platform
131 * ID stored in the structure are unique for each rank (even if a device
132 * is shared by multiple ranks).
134 * TODO: We should also do CPU hardware detection only once on each
135 * physical node and broadcast it, instead of do it on every MPI rank.
138 /* A split of MPI_COMM_WORLD over physical nodes is only required here,
139 * so we create and destroy it locally.
141 MPI_Comm_rank(MPI_COMM_WORLD, &rank_world);
142 MPI_Comm_split(MPI_COMM_WORLD, gmx_physicalnode_id_hash(),
143 rank_world, &physicalnode_comm);
144 MPI_Comm_rank(physicalnode_comm, &rank_local);
145 GMX_UNUSED_VALUE(cr);
147 /* Here there should be only one process, check this */
148 GMX_RELEASE_ASSERT(cr->nnodes == 1 && cr->sim_nodeid == 0, "Only a single (master) process should execute here");
153 /* With CUDA detect only on one rank per host, with OpenCL need do
154 * the detection on all PP ranks */
155 bool isOpenclPpRank = ((GMX_GPU == GMX_GPU_OPENCL) && thisRankHasDuty(cr, DUTY_PP));
157 if (rank_local == 0 || isOpenclPpRank)
159 char detection_error[STRLEN] = "", sbuf[STRLEN];
161 if (detect_gpus(&hwinfo_g->gpu_info, detection_error) != 0)
163 if (detection_error[0] != '\0')
165 sprintf(sbuf, ":\n %s\n", detection_error);
171 GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
172 "NOTE: Error occurred during GPU detection%s"
173 " Can not use GPU acceleration, will fall back to CPU kernels.",
181 /* Broadcast the GPU info to the other ranks within this node */
182 MPI_Bcast(&hwinfo_g->gpu_info.n_dev, 1, MPI_INT, 0, physicalnode_comm);
184 if (hwinfo_g->gpu_info.n_dev > 0)
188 dev_size = hwinfo_g->gpu_info.n_dev*sizeof_gpu_dev_info();
192 hwinfo_g->gpu_info.gpu_dev =
193 (struct gmx_device_info_t *)malloc(dev_size);
195 MPI_Bcast(hwinfo_g->gpu_info.gpu_dev, dev_size, MPI_BYTE,
196 0, physicalnode_comm);
197 MPI_Bcast(&hwinfo_g->gpu_info.n_dev_compatible, 1, MPI_INT,
198 0, physicalnode_comm);
202 MPI_Comm_free(&physicalnode_comm);
206 //! Reduce the locally collected \p hwinfo_g over MPI ranks
207 static void gmx_collect_hardware_mpi(const gmx::CpuInfo &cpuInfo)
209 const int ncore = hwinfo_g->hardwareTopology->numberOfCores();
212 int nrank, rank, nhwthread, ngpu, i;
216 rank_id = gmx_physicalnode_id_hash();
217 MPI_Comm_rank(MPI_COMM_WORLD, &rank);
218 MPI_Comm_size(MPI_COMM_WORLD, &nrank);
219 nhwthread = hwinfo_g->nthreads_hw_avail;
220 ngpu = hwinfo_g->gpu_info.n_dev_compatible;
221 /* Create a unique hash of the GPU type(s) in this node */
223 /* Here it might be better to only loop over the compatible GPU, but we
224 * don't have that information available and it would also require
225 * removing the device ID from the device info string.
227 for (i = 0; i < hwinfo_g->gpu_info.n_dev; i++)
231 /* Since the device ID is incorporated in the hash, the order of
232 * the GPUs affects the hash. Also two identical GPUs won't give
233 * a gpu_hash of zero after XORing.
235 get_gpu_device_info_string(stmp, hwinfo_g->gpu_info, i);
236 gpu_hash ^= gmx_string_fullhash_func(stmp, gmx_string_hash_init);
243 MPI_Allreduce(buf, all, nrank, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
246 int nnode0, ncore0, nhwthread0, ngpu0, r;
253 for (r = 0; r < nrank; r++)
255 if (all[r] == rank_id)
257 if (!bFound && r == rank)
259 /* We are the first rank in this physical node */
262 nhwthread0 = nhwthread;
272 int sum[4], maxmin[10];
277 /* Sum values from only intra-rank 0 so we get the sum over all nodes */
283 MPI_Allreduce(buf, sum, 4, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
289 /* Store + and - values for all ranks,
290 * so we can get max+min with one MPI call.
295 buf[3] = static_cast<int>(gmx::simdSuggested(cpuInfo));
303 MPI_Allreduce(buf, maxmin, 10, MPI_INT, MPI_MAX, MPI_COMM_WORLD);
306 hwinfo_g->nphysicalnode = sum[0];
307 hwinfo_g->ncore_tot = sum[1];
308 hwinfo_g->ncore_min = -maxmin[5];
309 hwinfo_g->ncore_max = maxmin[0];
310 hwinfo_g->nhwthread_tot = sum[2];
311 hwinfo_g->nhwthread_min = -maxmin[6];
312 hwinfo_g->nhwthread_max = maxmin[1];
313 hwinfo_g->ngpu_compatible_tot = sum[3];
314 hwinfo_g->ngpu_compatible_min = -maxmin[7];
315 hwinfo_g->ngpu_compatible_max = maxmin[2];
316 hwinfo_g->simd_suggest_min = -maxmin[8];
317 hwinfo_g->simd_suggest_max = maxmin[3];
318 hwinfo_g->bIdenticalGPUs = (maxmin[4] == -maxmin[9]);
320 /* All ranks use the same pointer, protected by a mutex in the caller */
321 hwinfo_g->nphysicalnode = 1;
322 hwinfo_g->ncore_tot = ncore;
323 hwinfo_g->ncore_min = ncore;
324 hwinfo_g->ncore_max = ncore;
325 hwinfo_g->nhwthread_tot = hwinfo_g->nthreads_hw_avail;
326 hwinfo_g->nhwthread_min = hwinfo_g->nthreads_hw_avail;
327 hwinfo_g->nhwthread_max = hwinfo_g->nthreads_hw_avail;
328 hwinfo_g->ngpu_compatible_tot = hwinfo_g->gpu_info.n_dev_compatible;
329 hwinfo_g->ngpu_compatible_min = hwinfo_g->gpu_info.n_dev_compatible;
330 hwinfo_g->ngpu_compatible_max = hwinfo_g->gpu_info.n_dev_compatible;
331 hwinfo_g->simd_suggest_min = static_cast<int>(simdSuggested(cpuInfo));
332 hwinfo_g->simd_suggest_max = static_cast<int>(simdSuggested(cpuInfo));
333 hwinfo_g->bIdenticalGPUs = TRUE;
337 /*! \brief Utility that does dummy computing for max 2 seconds to spin up cores
339 * This routine will check the number of cores configured and online
340 * (using sysconf), and the spins doing dummy compute operations for up to
341 * 2 seconds, or until all cores have come online. This can be used prior to
342 * hardware detection for platforms that take unused processors offline.
344 * This routine will not throw exceptions.
347 spinUpCore() noexcept
349 #if defined(HAVE_SYSCONF) && defined(_SC_NPROCESSORS_CONF) && defined(_SC_NPROCESSORS_ONLN)
351 int countConfigured = sysconf(_SC_NPROCESSORS_CONF); // noexcept
352 auto start = std::chrono::steady_clock::now(); // noexcept
354 while (sysconf(_SC_NPROCESSORS_ONLN) < countConfigured &&
355 std::chrono::steady_clock::now() - start < std::chrono::seconds(2))
357 for (int i = 1; i < 10000; i++)
365 printf("This cannot happen, but prevents loop from being optimized away.");
370 /*! \brief Prepare the system before hardware topology detection
372 * This routine should perform any actions we want to put the system in a state
373 * where we want it to be before detecting the hardware topology. For most
374 * processors there is nothing to do, but some architectures (in particular ARM)
375 * have support for taking configured cores offline, which will make them disappear
376 * from the online processor count.
378 * This routine checks if there is a mismatch between the number of cores
379 * configured and online, and in that case we issue a small workload that
380 * attempts to wake sleeping cores before doing the actual detection.
382 * This type of mismatch can also occur for x86 or PowerPC on Linux, if SMT has only
383 * been disabled in the kernel (rather than bios). Since those cores will never
384 * come online automatically, we currently skip this test for x86 & PowerPC to
385 * avoid wasting 2 seconds. We also skip the test if there is no thread support.
387 * \note Cores will sleep relatively quickly again, so it's important to issue
388 * the real detection code directly after this routine.
391 hardwareTopologyPrepareDetection()
393 #if defined(HAVE_SYSCONF) && defined(_SC_NPROCESSORS_CONF) && \
394 (defined(THREAD_PTHREADS) || defined(THREAD_WINDOWS))
396 // Modify this conditional when/if x86 or PowerPC starts to sleep some cores
397 if (c_architecture != Architecture::X86 &&
398 c_architecture != Architecture::PowerPC)
400 int countConfigured = sysconf(_SC_NPROCESSORS_CONF);
401 std::vector<std::thread> workThreads(countConfigured);
403 for (auto &t : workThreads)
405 t = std::thread(spinUpCore);
408 for (auto &t : workThreads)
416 /*! \brief Sanity check hardware topology and print some notes to log
418 * \param mdlog Logger.
419 * \param hardwareTopology Reference to hardwareTopology object.
422 hardwareTopologyDoubleCheckDetection(const gmx::MDLogger gmx_unused &mdlog,
423 const gmx::HardwareTopology gmx_unused &hardwareTopology)
425 #if defined HAVE_SYSCONF && defined(_SC_NPROCESSORS_CONF)
426 if (hardwareTopology.supportLevel() < gmx::HardwareTopology::SupportLevel::LogicalProcessorCount)
431 int countFromDetection = hardwareTopology.machine().logicalProcessorCount;
432 int countConfigured = sysconf(_SC_NPROCESSORS_CONF);
434 /* BIOS, kernel or user actions can take physical processors
435 * offline. We already cater for the some of the cases inside the hardwareToplogy
436 * by trying to spin up cores just before we detect, but there could be other
437 * cases where it is worthwhile to hint that there might be more resources available.
439 if (countConfigured >= 0 && countConfigured != countFromDetection)
442 appendTextFormatted("Note: %d CPUs configured, but only %d were detected to be online.\n", countConfigured, countFromDetection);
444 if (c_architecture == Architecture::X86 &&
445 countConfigured == 2*countFromDetection)
448 appendText(" X86 Hyperthreading is likely disabled; enable it for better performance.");
450 // For PowerPC (likely Power8) it is possible to set SMT to either 2,4, or 8-way hardware threads.
451 // We only warn if it is completely disabled since default performance drops with SMT8.
452 if (c_architecture == Architecture::PowerPC &&
453 countConfigured == 8*countFromDetection)
456 appendText(" PowerPC SMT is likely disabled; enable SMT2/SMT4 for better performance.");
462 gmx_hw_info_t *gmx_detect_hardware(const gmx::MDLogger &mdlog, const t_commrec *cr)
466 /* make sure no one else is doing the same thing */
467 ret = tMPI_Thread_mutex_lock(&hw_info_lock);
470 gmx_fatal(FARGS, "Error locking hwinfo mutex: %s", strerror(errno));
473 /* only initialize the hwinfo structure if it is not already initalized */
476 hwinfo_g = compat::make_unique<gmx_hw_info_t>();
478 hwinfo_g->cpuInfo = new gmx::CpuInfo(gmx::CpuInfo::detect());
480 hardwareTopologyPrepareDetection();
481 hwinfo_g->hardwareTopology = new gmx::HardwareTopology(gmx::HardwareTopology::detect());
483 // If we detected the topology on this system, double-check that it makes sense
484 if (hwinfo_g->hardwareTopology->isThisSystem())
486 hardwareTopologyDoubleCheckDetection(mdlog, *(hwinfo_g->hardwareTopology));
489 // TODO: Get rid of this altogether.
490 hwinfo_g->nthreads_hw_avail = hwinfo_g->hardwareTopology->machine().logicalProcessorCount;
493 hwinfo_g->gpu_info.n_dev = 0;
494 hwinfo_g->gpu_info.n_dev_compatible = 0;
495 hwinfo_g->gpu_info.gpu_dev = nullptr;
497 gmx_detect_gpus(mdlog, cr);
498 gmx_collect_hardware_mpi(*hwinfo_g->cpuInfo);
499 hwinfo_g->compatibleGpus = getCompatibleGpus(hwinfo_g->gpu_info);
501 /* increase the reference counter */
504 ret = tMPI_Thread_mutex_unlock(&hw_info_lock);
507 gmx_fatal(FARGS, "Error unlocking hwinfo mutex: %s", strerror(errno));
510 return hwinfo_g.get();
513 bool compatibleGpusFound(const gmx_gpu_info_t &gpu_info)
515 return gpu_info.n_dev_compatible > 0;
518 void gmx_hardware_info_free()
522 ret = tMPI_Thread_mutex_lock(&hw_info_lock);
525 gmx_fatal(FARGS, "Error locking hwinfo mutex: %s", strerror(errno));
528 /* decrease the reference counter */
534 gmx_incons("n_hwinfo < 0");
539 delete hwinfo_g->cpuInfo;
540 delete hwinfo_g->hardwareTopology;
541 free_gpu_info(&hwinfo_g->gpu_info);
545 ret = tMPI_Thread_mutex_unlock(&hw_info_lock);
548 gmx_fatal(FARGS, "Error unlocking hwinfo mutex: %s", strerror(errno));