2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015,2016, 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/gpu_utils/gpu_utils.h"
52 #include "gromacs/hardware/cpuinfo.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/logger.h"
64 #include "gromacs/utility/mutex.h"
65 #include "gromacs/utility/physicalnodecommunicator.h"
67 #include "architecture.h"
70 # include <unistd.h> // sysconf()
73 gmx_hw_info_t::gmx_hw_info_t(std::unique_ptr
<gmx::CpuInfo
> cpuInfo
,
74 std::unique_ptr
<gmx::HardwareTopology
> hardwareTopology
) :
75 cpuInfo(std::move(cpuInfo
)),
76 hardwareTopology(std::move(hardwareTopology
))
80 gmx_hw_info_t::~gmx_hw_info_t()
82 free_gpu_info(&gpu_info
);
88 //! Convenience macro to help us avoid ifdefs each time we use sysconf
89 #if !defined(_SC_NPROCESSORS_ONLN) && defined(_SC_NPROC_ONLN)
90 # define _SC_NPROCESSORS_ONLN _SC_NPROC_ONLN
93 //! Convenience macro to help us avoid ifdefs each time we use sysconf
94 #if !defined(_SC_NPROCESSORS_CONF) && defined(_SC_NPROC_CONF)
95 # define _SC_NPROCESSORS_CONF _SC_NPROC_CONF
98 /*! \brief Information about the hardware of all nodes (common to all threads in this process).
100 * This information is constructed only when required, but thereafter
101 * its lifetime is that of the whole process, potentially across
102 * multiple successive simulation parts. It's wise to ensure that only
103 * one thread can create the information, but thereafter they can all
104 * read it without e.g. needing a std::shared_ptr to ensure its
105 * lifetime exceeds that of the thread. */
106 static std::unique_ptr
<gmx_hw_info_t
> g_hardwareInfo
;
107 //! A mutex to protect the hwinfo structure
108 static Mutex g_hardwareInfoMutex
;
110 //! Detect GPUs, if that makes sense to attempt.
111 static void gmx_detect_gpus(const gmx::MDLogger
& mdlog
,
112 const PhysicalNodeCommunicator
& physicalNodeComm
,
113 compat::not_null
<gmx_hw_info_t
*> hardwareInfo
)
115 hardwareInfo
->gpu_info
.bDetectGPUs
= canPerformGpuDetection();
117 if (!hardwareInfo
->gpu_info
.bDetectGPUs
)
122 bool isMasterRankOfPhysicalNode
= true;
124 isMasterRankOfPhysicalNode
= (physicalNodeComm
.rank_
== 0);
126 // We choose to run the detection only once with thread-MPI and
127 // use a mutex to enforce it.
128 GMX_UNUSED_VALUE(physicalNodeComm
);
129 isMasterRankOfPhysicalNode
= true;
132 /* The OpenCL support requires us to run detection on all ranks.
133 * With CUDA we don't need to, and prefer to detect on one rank
134 * and send the information to the other ranks over MPI. */
135 bool allRanksMustDetectGpus
= (GMX_GPU
== GMX_GPU_OPENCL
);
136 bool gpusCanBeDetected
= false;
137 if (isMasterRankOfPhysicalNode
|| allRanksMustDetectGpus
)
139 std::string errorMessage
;
140 gpusCanBeDetected
= isGpuDetectionFunctional(&errorMessage
);
141 if (!gpusCanBeDetected
)
145 .appendTextFormatted(
146 "NOTE: Detection of GPUs failed. The API reported:\n"
148 " GROMACS cannot run tasks on a GPU.",
149 errorMessage
.c_str());
153 if (gpusCanBeDetected
)
155 findGpus(&hardwareInfo
->gpu_info
);
156 // No need to tell the user anything at this point, they get a
157 // hardware report later.
161 if (!allRanksMustDetectGpus
)
163 /* Broadcast the GPU info to the other ranks within this node */
164 MPI_Bcast(&hardwareInfo
->gpu_info
.n_dev
, 1, MPI_INT
, 0, physicalNodeComm
.comm_
);
166 if (hardwareInfo
->gpu_info
.n_dev
> 0)
170 dev_size
= hardwareInfo
->gpu_info
.n_dev
* sizeof_gpu_dev_info();
172 if (!isMasterRankOfPhysicalNode
)
174 hardwareInfo
->gpu_info
.gpu_dev
= (struct gmx_device_info_t
*)malloc(dev_size
);
176 MPI_Bcast(hardwareInfo
->gpu_info
.gpu_dev
, dev_size
, MPI_BYTE
, 0, physicalNodeComm
.comm_
);
177 MPI_Bcast(&hardwareInfo
->gpu_info
.n_dev_compatible
, 1, MPI_INT
, 0, physicalNodeComm
.comm_
);
183 //! Reduce the locally collected \p hardwareInfo over MPI ranks
184 static void gmx_collect_hardware_mpi(const gmx::CpuInfo
& cpuInfo
,
185 const PhysicalNodeCommunicator
& physicalNodeComm
,
186 compat::not_null
<gmx_hw_info_t
*> hardwareInfo
)
188 const int ncore
= hardwareInfo
->hardwareTopology
->numberOfCores();
189 /* Zen1 is assumed for:
190 * - family=23 with the below listed models;
193 const bool cpuIsAmdZen1
= ((cpuInfo
.vendor() == CpuInfo::Vendor::Amd
&& cpuInfo
.family() == 23
194 && (cpuInfo
.model() == 1 || cpuInfo
.model() == 17
195 || cpuInfo
.model() == 8 || cpuInfo
.model() == 24))
196 || cpuInfo
.vendor() == CpuInfo::Vendor::Hygon
);
198 int nhwthread
, ngpu
, i
;
201 nhwthread
= hardwareInfo
->nthreads_hw_avail
;
202 ngpu
= hardwareInfo
->gpu_info
.n_dev_compatible
;
203 /* Create a unique hash of the GPU type(s) in this node */
205 /* Here it might be better to only loop over the compatible GPU, but we
206 * don't have that information available and it would also require
207 * removing the device ID from the device info string.
209 for (i
= 0; i
< hardwareInfo
->gpu_info
.n_dev
; i
++)
213 /* Since the device ID is incorporated in the hash, the order of
214 * the GPUs affects the hash. Also two identical GPUs won't give
215 * a gpu_hash of zero after XORing.
217 get_gpu_device_info_string(stmp
, hardwareInfo
->gpu_info
, i
);
218 gpu_hash
^= gmx_string_fullhash_func(stmp
, gmx_string_hash_init
);
221 constexpr int numElementsCounts
= 4;
222 std::array
<int, numElementsCounts
> countsReduced
;
224 std::array
<int, numElementsCounts
> countsLocal
= { { 0 } };
225 // Organize to sum values from only one rank within each node,
226 // so we get the sum over all nodes.
227 bool isMasterRankOfPhysicalNode
= (physicalNodeComm
.rank_
== 0);
228 if (isMasterRankOfPhysicalNode
)
231 countsLocal
[1] = ncore
;
232 countsLocal
[2] = nhwthread
;
233 countsLocal
[3] = ngpu
;
236 MPI_Allreduce(countsLocal
.data(), countsReduced
.data(), countsLocal
.size(), MPI_INT
,
237 MPI_SUM
, MPI_COMM_WORLD
);
240 constexpr int numElementsMax
= 11;
241 std::array
<int, numElementsMax
> maxMinReduced
;
243 std::array
<int, numElementsMax
> maxMinLocal
;
244 /* Store + and - values for all ranks,
245 * so we can get max+min with one MPI call.
247 maxMinLocal
[0] = ncore
;
248 maxMinLocal
[1] = nhwthread
;
249 maxMinLocal
[2] = ngpu
;
250 maxMinLocal
[3] = static_cast<int>(gmx::simdSuggested(cpuInfo
));
251 maxMinLocal
[4] = gpu_hash
;
252 maxMinLocal
[5] = -maxMinLocal
[0];
253 maxMinLocal
[6] = -maxMinLocal
[1];
254 maxMinLocal
[7] = -maxMinLocal
[2];
255 maxMinLocal
[8] = -maxMinLocal
[3];
256 maxMinLocal
[9] = -maxMinLocal
[4];
257 maxMinLocal
[10] = (cpuIsAmdZen1
? 1 : 0);
259 MPI_Allreduce(maxMinLocal
.data(), maxMinReduced
.data(), maxMinLocal
.size(), MPI_INT
,
260 MPI_MAX
, MPI_COMM_WORLD
);
263 hardwareInfo
->nphysicalnode
= countsReduced
[0];
264 hardwareInfo
->ncore_tot
= countsReduced
[1];
265 hardwareInfo
->ncore_min
= -maxMinReduced
[5];
266 hardwareInfo
->ncore_max
= maxMinReduced
[0];
267 hardwareInfo
->nhwthread_tot
= countsReduced
[2];
268 hardwareInfo
->nhwthread_min
= -maxMinReduced
[6];
269 hardwareInfo
->nhwthread_max
= maxMinReduced
[1];
270 hardwareInfo
->ngpu_compatible_tot
= countsReduced
[3];
271 hardwareInfo
->ngpu_compatible_min
= -maxMinReduced
[7];
272 hardwareInfo
->ngpu_compatible_max
= maxMinReduced
[2];
273 hardwareInfo
->simd_suggest_min
= -maxMinReduced
[8];
274 hardwareInfo
->simd_suggest_max
= maxMinReduced
[3];
275 hardwareInfo
->bIdenticalGPUs
= (maxMinReduced
[4] == -maxMinReduced
[9]);
276 hardwareInfo
->haveAmdZen1Cpu
= (maxMinReduced
[10] > 0);
278 /* All ranks use the same pointer, protected by a mutex in the caller */
279 hardwareInfo
->nphysicalnode
= 1;
280 hardwareInfo
->ncore_tot
= ncore
;
281 hardwareInfo
->ncore_min
= ncore
;
282 hardwareInfo
->ncore_max
= ncore
;
283 hardwareInfo
->nhwthread_tot
= hardwareInfo
->nthreads_hw_avail
;
284 hardwareInfo
->nhwthread_min
= hardwareInfo
->nthreads_hw_avail
;
285 hardwareInfo
->nhwthread_max
= hardwareInfo
->nthreads_hw_avail
;
286 hardwareInfo
->ngpu_compatible_tot
= hardwareInfo
->gpu_info
.n_dev_compatible
;
287 hardwareInfo
->ngpu_compatible_min
= hardwareInfo
->gpu_info
.n_dev_compatible
;
288 hardwareInfo
->ngpu_compatible_max
= hardwareInfo
->gpu_info
.n_dev_compatible
;
289 hardwareInfo
->simd_suggest_min
= static_cast<int>(simdSuggested(cpuInfo
));
290 hardwareInfo
->simd_suggest_max
= static_cast<int>(simdSuggested(cpuInfo
));
291 hardwareInfo
->bIdenticalGPUs
= TRUE
;
292 hardwareInfo
->haveAmdZen1Cpu
= cpuIsAmdZen1
;
293 GMX_UNUSED_VALUE(physicalNodeComm
);
297 /*! \brief Utility that does dummy computing for max 2 seconds to spin up cores
299 * This routine will check the number of cores configured and online
300 * (using sysconf), and the spins doing dummy compute operations for up to
301 * 2 seconds, or until all cores have come online. This can be used prior to
302 * hardware detection for platforms that take unused processors offline.
304 * This routine will not throw exceptions. In principle it should be
305 * declared noexcept, but at least icc 19.1 and 21-beta08 with the
306 * libstdc++-7.5 has difficulty implementing a std::vector of
307 * std::thread started with this function when declared noexcept. It
308 * is a known compiler bug that should be fixed after 19.1.
309 * Fortunately, this function is not performance sensitive,
310 * and only runs on platforms other than x86 and POWER (ie ARM),
311 * so the possible overhead introduced by omitting noexcept is not
314 static void spinUpCore()
316 #if defined(HAVE_SYSCONF) && defined(_SC_NPROCESSORS_CONF) && defined(_SC_NPROCESSORS_ONLN)
318 int countConfigured
= sysconf(_SC_NPROCESSORS_CONF
); // noexcept
319 auto start
= std::chrono::steady_clock::now(); // noexcept
321 while (sysconf(_SC_NPROCESSORS_ONLN
) < countConfigured
322 && std::chrono::steady_clock::now() - start
< std::chrono::seconds(2))
324 for (int i
= 1; i
< 10000; i
++)
332 printf("This cannot happen, but prevents loop from being optimized away.");
337 /*! \brief Prepare the system before hardware topology detection
339 * This routine should perform any actions we want to put the system in a state
340 * where we want it to be before detecting the hardware topology. For most
341 * processors there is nothing to do, but some architectures (in particular ARM)
342 * have support for taking configured cores offline, which will make them disappear
343 * from the online processor count.
345 * This routine checks if there is a mismatch between the number of cores
346 * configured and online, and in that case we issue a small workload that
347 * attempts to wake sleeping cores before doing the actual detection.
349 * This type of mismatch can also occur for x86 or PowerPC on Linux, if SMT has only
350 * been disabled in the kernel (rather than bios). Since those cores will never
351 * come online automatically, we currently skip this test for x86 & PowerPC to
352 * avoid wasting 2 seconds. We also skip the test if there is no thread support.
354 * \note Cores will sleep relatively quickly again, so it's important to issue
355 * the real detection code directly after this routine.
357 static void hardwareTopologyPrepareDetection()
359 #if defined(HAVE_SYSCONF) && defined(_SC_NPROCESSORS_CONF) \
360 && (defined(THREAD_PTHREADS) || defined(THREAD_WINDOWS))
362 // Modify this conditional when/if x86 or PowerPC starts to sleep some cores
363 if (c_architecture
!= Architecture::X86
&& c_architecture
!= Architecture::PowerPC
)
365 int countConfigured
= sysconf(_SC_NPROCESSORS_CONF
);
366 std::vector
<std::thread
> workThreads(countConfigured
);
368 for (auto& t
: workThreads
)
370 t
= std::thread(spinUpCore
);
373 for (auto& t
: workThreads
)
381 /*! \brief Sanity check hardware topology and print some notes to log
383 * \param mdlog Logger.
384 * \param hardwareTopology Reference to hardwareTopology object.
386 static void hardwareTopologyDoubleCheckDetection(const gmx::MDLogger gmx_unused
& mdlog
,
387 const gmx::HardwareTopology gmx_unused
& hardwareTopology
)
389 #if defined HAVE_SYSCONF && defined(_SC_NPROCESSORS_CONF)
390 if (hardwareTopology
.supportLevel() < gmx::HardwareTopology::SupportLevel::LogicalProcessorCount
)
395 int countFromDetection
= hardwareTopology
.machine().logicalProcessorCount
;
396 int countConfigured
= sysconf(_SC_NPROCESSORS_CONF
);
398 /* BIOS, kernel or user actions can take physical processors
399 * offline. We already cater for the some of the cases inside the hardwareToplogy
400 * by trying to spin up cores just before we detect, but there could be other
401 * cases where it is worthwhile to hint that there might be more resources available.
403 if (countConfigured
>= 0 && countConfigured
!= countFromDetection
)
406 .appendTextFormatted(
407 "Note: %d CPUs configured, but only %d were detected to be online.\n",
408 countConfigured
, countFromDetection
);
410 if (c_architecture
== Architecture::X86
&& countConfigured
== 2 * countFromDetection
)
414 " X86 Hyperthreading is likely disabled; enable it for better "
417 // For PowerPC (likely Power8) it is possible to set SMT to either 2,4, or 8-way hardware threads.
418 // We only warn if it is completely disabled since default performance drops with SMT8.
419 if (c_architecture
== Architecture::PowerPC
&& countConfigured
== 8 * countFromDetection
)
423 " PowerPC SMT is likely disabled; enable SMT2/SMT4 for better "
430 gmx_hw_info_t
* gmx_detect_hardware(const gmx::MDLogger
& mdlog
, const PhysicalNodeCommunicator
& physicalNodeComm
)
432 // By construction, only one thread ever runs hardware detection,
433 // but we may as well prevent issues arising if that would change.
434 // Taking the lock early ensures that exactly one thread can
435 // attempt to construct g_hardwareInfo.
436 lock_guard
<Mutex
> lock(g_hardwareInfoMutex
);
438 // If we already have the information, just return a handle to it.
439 if (g_hardwareInfo
!= nullptr)
441 return g_hardwareInfo
.get();
444 // Make the new hardwareInfo in a temporary.
445 hardwareTopologyPrepareDetection();
447 // TODO: We should also do CPU hardware detection only once on each
448 // physical node and broadcast it, instead of doing it on every MPI rank.
449 auto hardwareInfo
= std::make_unique
<gmx_hw_info_t
>(
450 std::make_unique
<CpuInfo
>(CpuInfo::detect()),
451 std::make_unique
<HardwareTopology
>(HardwareTopology::detect()));
453 // If we detected the topology on this system, double-check that it makes sense
454 if (hardwareInfo
->hardwareTopology
->isThisSystem())
456 hardwareTopologyDoubleCheckDetection(mdlog
, *hardwareInfo
->hardwareTopology
);
459 // TODO: Get rid of this altogether.
460 hardwareInfo
->nthreads_hw_avail
= hardwareInfo
->hardwareTopology
->machine().logicalProcessorCount
;
463 hardwareInfo
->gpu_info
.n_dev
= 0;
464 hardwareInfo
->gpu_info
.n_dev_compatible
= 0;
465 hardwareInfo
->gpu_info
.gpu_dev
= nullptr;
467 gmx_detect_gpus(mdlog
, physicalNodeComm
, compat::make_not_null(hardwareInfo
));
468 gmx_collect_hardware_mpi(*hardwareInfo
->cpuInfo
, physicalNodeComm
, compat::make_not_null(hardwareInfo
));
470 // Now that the temporary is fully constructed, swap it to become
472 g_hardwareInfo
.swap(hardwareInfo
);
474 return g_hardwareInfo
.get();
477 bool compatibleGpusFound(const gmx_gpu_info_t
& gpu_info
)
479 return gpu_info
.n_dev_compatible
> 0;