Merge pull request #270 from ls1mardyn/updateAutoPasCombinedTuning

Update AutoPas to Combined Tuning

.github/workflows/ls1_test.yml

@@ -91,7 +91,7 @@ jobs:
                 -DENABLE_UNIT_TESTS=ON \
               ..
         
-          cmake --build . --parallel 2
+          cmake --build . --parallel 1
 
           cd ..
 
@@ -151,7 +151,7 @@ jobs:
                 -DENABLE_MPI=$mpi_enabled \
               ..
         
-          cmake --build . --parallel 2
+          cmake --build . --parallel 1
 
           #as example list of new version is used, also the example files of new version should be used
           #therefore, go back to new version

cmake/modules/autopas.cmake

@@ -23,8 +23,8 @@ if (ENABLE_AUTOPAS)
     set(AUTOPAS_USE_VECTORIZATION ${USE_VECTORIZATION} CACHE BOOL "Set via USE_VECTORIZATION" FORCE)
     set(AUTOPAS_VECTOR_INSTRUCTIONS ${VECTOR_INSTRUCTIONS} CACHE STRING "Set via VECTOR_INSTRUCTIONS_OPTIONS" FORCE)
 
-    # Merge of fixSpacialLocksResize 2023-05-08; This includes the rewrites of remainder traversal and iterators.
-    set(AUTOPAS_TAG e47c9257bb8d0c6ab680d02823b1e8749b25369a CACHE STRING "AutoPas Git tag or commit id to use")
+    # Merge of #830 (fixRegionIterators); includes #787 (combined tuning)
+    set(AUTOPAS_TAG faef573 CACHE STRING "AutoPas Git tag or commit id to use")
 
     FetchContent_Declare(
             autopasfetch
@@ -41,7 +41,8 @@ if (ENABLE_AUTOPAS)
         add_subdirectory(${autopasfetch_SOURCE_DIR} ${autopasfetch_BINARY_DIR} EXCLUDE_FROM_ALL)
     endif ()
 
-    set(AUTOPAS_LIB "autopas")
+    # Linking the md lib also links AutoPas
+    set(AUTOPAS_LIB "molecularDynamicsLibrary")
 else ()
     message(STATUS "Not using AutoPas.")
     set(AUTOPAS_LIB "")

examples/Adsorption/Generator/adsorption_autopas_tuning.xml

@@ -87,7 +87,6 @@
         <allowedTraversals>c01, c04, c08, c18, sli</allowedTraversals>
         <allowedContainers>linkedCells</allowedContainers>
         <selectorStrategy>fastestMedian</selectorStrategy>
-        <tuningStrategy>FullSearch</tuningStrategy><!--FullSearch, BayesianSearch, ...-->
         <dataLayouts>SoA</dataLayouts>
         <newton3>enabled</newton3>
         <tuningInterval>1000</tuningInterval>

examples/Argon/200K_18mol_l/config_autopas_allOptions.xml

@@ -44,16 +44,18 @@
       </parallelisation>
       -->
       <datastructure type="AutoPas">
+        <loglevel>off</loglevel><!--Level for the AutoPas logger-->
         <allowedTraversals>ds_sequential, lc_sliced, lc_sliced_balanced, lc_sliced_c02, lc_c01, lc_c01_combined_SoA, lc_c01_cuda, lc_c04, lc_c04_HCP, lc_c04_combined_SoA, lc_c08, lc_c18, vcc_cluster_iteration_cuda, vcl_cluster_iteration, vcl_c06, vcl_c01_balanced, vl_list_iteration, vlc_c01, vlc_c18, vlc_sliced, vlc_sliced_balanced, vlc_sliced_c02, vvl_as_built</allowedTraversals>
         <allowedContainers>DirectSum,LinkedCells, VerletLists, VerletListsCells, VerletClusterLists,VarVerletListsAsBuild,VerletClusterCells</allowedContainers>
         <selectorStrategy>fastestMean</selectorStrategy><!--fastestMedian,fastestAbs-->
-        <tuningStrategy>FullSearch</tuningStrategy><!--FullSearch, BayesianSearch, ...-->
+        <tuningStrategies>predictive,rulebased</tuningStrategies><!--Predictive, RuleBased, BayesianSearch, ...; Leave empty for full search-->
         <dataLayouts>AoS, SoA</dataLayouts>
         <newton3>enabled, disabled</newton3>
         <tuningAcquisitionFunction>lower-confidence-bound</tuningAcquisitionFunction> <!--only relevant for BayesianSearch-->
         <maxEvidence>20</maxEvidence>
         <tuningInterval>1000</tuningInterval>
         <tuningSamples>10</tuningSamples>
+        <ruleFile>AutoPas/examples/md-flexible/input/tuningRules.rule</ruleFile><!--check the AutoPas repo for examples-->
         <rebuildFrequency>10</rebuildFrequency>
         <skin>0.5</skin>
       </datastructure>

examples/Argon/200K_18mol_l/config_autopas_aos.xml

@@ -35,6 +35,7 @@
         <allowedTraversals>lc_c01, lc_c08, lc_c18, lc_sli, ds_sequential</allowedTraversals>
         <allowedContainers>linkedCells, verletLists, directsum</allowedContainers>
         <selectorStrategy>fastestMedian</selectorStrategy>
+        <tuningMetric>energy</tuningMetric>
         <dataLayouts>AoS</dataLayouts>
         <newton3>enabled, disabled</newton3>
         <tuningInterval>1000</tuningInterval>

examples/Argon/200K_18mol_l/config_autopas_lc_ALL.xml

@@ -38,7 +38,9 @@
                 </loadBalancer>
             </parallelisation>
             <datastructure type="AutoPas">
+                <!-- <logLevel>info</logLevel> -->
                 <allowedContainers>linkedCells</allowedContainers>
+                <tuningStrategies>predictive,ruleBased</tuningStrategies>
                 <tuningInterval>10000</tuningInterval>
                 <tuningSamples>10</tuningSamples>
                 <rebuildFrequency>10</rebuildFrequency>

examples/all-options.xml

@@ -263,8 +263,11 @@
         <!-- specify what metric the selector should use to determine the optimal traversal -->
         <!-- possible values: fastetAbsolute, fastestMean, fastestMedian -->
         <selectorStrategy>fastestAbsolute</selectorStrategy>
-        <!-- set the auto tuning algorithm -->
-        <tuningStrategy>fullSearch</tuningStrategy>
+        <!-- set the tuning auto tuning pipeline -->
+        <!-- Leave this empty for an exhaustive full search or add any number of strategies that are applied in the stated order -->
+        <tuningStrategies>ruleBasedTuning, predictiveTuning</tuningStrategies>
+        <!-- autopas optimizes the simulation based on tuning metric: time, energy, default option is time -->
+        <tuningMetric>energy</tuningMetric>
         <!-- Data layouts to be used: AoS, SoA -->
         <dataLayouts>SoA</dataLayouts>
         <!-- Allowed choices for newton 3 optimization: enabled, disabled -->

src/Domain.cpp

@@ -261,30 +261,27 @@ void Domain::calculateGlobalValues(
 			Log::global_log->debug() << "Selective thermostat will be applied to set " << thermit->first
 				<< " (beta_trans = " << this->_universalBTrans[thermit->first]
 				<< ", beta_rot = " << this->_universalBRot[thermit->first] << "!)" << std::endl;
-			int rot_dof;
 			const double limit_energy =  KINLIMIT_PER_T * Ti;
 
 			#if defined(_OPENMP)
 			#pragma omp parallel
 			#endif
 			{
-
-				double Utrans, Urot, limit_rot_energy, vcorr, Dcorr;
 				for (auto tM = particleContainer->iterator(ParticleIterator::ONLY_INNER_AND_BOUNDARY); tM.isValid(); ++tM) {
-					Utrans = tM->U_trans();
+					const auto Utrans = tM->U_trans();
 					if (Utrans > limit_energy) {
-						vcorr = sqrt(limit_energy / Utrans);
+						const auto vcorr = sqrt(limit_energy / Utrans);
 						Log::global_log->debug() << ": v(m" << tM->getID() << ") *= " << vcorr << std::endl;
 						tM->scale_v(vcorr);
 						tM->scale_F(vcorr);
 					}
 
-					rot_dof = tM->component()->getRotationalDegreesOfFreedom();
+					const auto rot_dof = tM->component()->getRotationalDegreesOfFreedom();
 					if (rot_dof > 0) {
-						limit_rot_energy = 3.0 * rot_dof * Ti;
-						Urot = tM->U_rot();
+						const auto limit_rot_energy = 3.0 * rot_dof * Ti;
+						const auto Urot = tM->U_rot();
 						if (Urot > limit_rot_energy) {
-							Dcorr = sqrt(limit_rot_energy / Urot);
+							const auto Dcorr = sqrt(limit_rot_energy / Urot);
 							Log::global_log->debug() << "D(m" << tM->getID() << ") *= " << Dcorr << std::endl;
 							tM->scale_D(Dcorr);
 							tM->scale_M(Dcorr);

src/molecules/AutoPasSimpleMolecule.h

@@ -6,7 +6,7 @@
 
 #pragma once
 
-#include <autopas/molecularDynamics/MoleculeLJ.h>
+#include <molecularDynamicsLibrary/MoleculeLJ.h>
 #include <autopas/utils/SoAType.h>
 #include <autopas/utils/inBox.h>
 
@@ -334,4 +334,4 @@ class AutoPasSimpleMolecule final : public MoleculeInterface, public autopas::Pa
 	static Quaternion _quaternion;
 };
 
-std::ostream& operator<<( std::ostream& os, const AutoPasSimpleMolecule& m );
+std::ostream& operator<<( std::ostream& os, const AutoPasSimpleMolecule& m );

src/parallel/CommunicationPartner.cpp

@@ -265,7 +265,7 @@ bool CommunicationPartner::iprobeCount(const MPI_Comm& comm, const MPI_Datatype&
 	return _countReceived;
 }
 bool CommunicationPartner::testRecv(ParticleContainer* moleculeContainer, bool removeRecvDuplicates, bool force) {
-		if (_countReceived and not _msgReceived) {
+	if (_countReceived and not _msgReceived) {
 		int flag = 1;
 		if (_countTested > 10) {
 			// some MPI (Intel, IBM) implementations can produce deadlocks using MPI_Test without any MPI_Wait

src/parallel/DomainDecompBase.cpp

@@ -260,7 +260,7 @@ void DomainDecompBase::handleDomainLeavingParticlesDirect(const HaloRegion& halo
 	};
 
 	if (moleculeContainer->isInvalidParticleReturner()) {
-		// move all particles that will be inserted now to the end of the container
+		// Shift and add all invalid particles that do belong in this halo region
 		auto removeBegin = std::partition(invalidParticles.begin(), invalidParticles.end(), [=](const Molecule& m) {
 			// if this is true, it will be put in the first part of the partition, if it is false, in the second.
 			return not m.inBox(haloRegion.rmin, haloRegion.rmax);

src/parallel/DomainDecompMPIBase.cpp

@@ -247,7 +247,7 @@ void DomainDecompMPIBase::assertDisjunctivity(ParticleContainer* moleculeContain
 			}
 		}
 		if (not isOk) {
-			Log::global_log->error() << "Aborting because of duplicated partices." << std::endl;
+			Log::global_log->error() << "Aborting because of duplicated particles." << std::endl;
 			MPI_Abort(_comm, 1);
 		}
 

src/parallel/NeighbourCommunicationScheme.cpp

@@ -299,6 +299,7 @@ void DirectNeighbourCommunicationScheme::finalizeExchangeMoleculesMPI(ParticleCo
 		removeRecvDuplicates |= (domainDecomp->getRank() == (*_neighbours)[0][i].getRank());
 	}
 
+	// local helper function to apply f to all real neighbours
 	auto forAllRealNeighbors = [&](auto&& f) {
 		for (auto& neighbor : (*_neighbours)[0]) {
 			if (not _useSequentialFallback or domainDecomp->getRank() != neighbor.getRank()) {