Block-Structured AMR Software Framework
Loading...
Searching...
No Matches
AMReX_DistributionMapping.H
Go to the documentation of this file.
1#ifndef BL_DISTRIBUTIONMAPPING_H
2#define BL_DISTRIBUTIONMAPPING_H
3#include <AMReX_Config.H>
4
5#include <AMReX.H>
6#include <AMReX_Array.H>
7#include <AMReX_Vector.H>
8#include <AMReX_Box.H>
9#include <AMReX_REAL.H>
11
12#include <map>
13#include <limits>
14#include <memory>
15#include <cstddef>
16#include <iosfwd>
17
18namespace amrex {
19
20class BoxArray;
21class MultiFab;
22template <typename T> class FabArray;
23template <typename T> class LayoutData;
24class FabArrayBase;
25
43{
44 public:
45
46 template <typename T> friend class FabArray;
47 friend class FabArrayBase;
48
51
52 struct Ref
53 {
55 Ref () = default;
56
57 explicit Ref (int len) : m_pmap(len) {}
58
59 explicit Ref (const Vector<int>& pmap) : m_pmap(pmap) {}
60
61 explicit Ref (Vector<int>&& pmap) noexcept : m_pmap(std::move(pmap)) {}
62
64
65 void clear () { m_pmap.clear(); m_index_array.clear(); m_ownership.clear(); }
66
69 std::vector<bool> m_ownership;
70 };
71
73 DistributionMapping () noexcept;
74
77
79 DistributionMapping& operator= (const DistributionMapping& rhs) = default;
80
82 DistributionMapping (DistributionMapping&& rhs) noexcept = default;
83
85 DistributionMapping& operator= (DistributionMapping&& rhs) noexcept = default;
86
88 ~DistributionMapping() noexcept = default;
89
93 explicit DistributionMapping (const Vector<int>& pmap);
94 explicit DistributionMapping (Vector<int>&& pmap) noexcept;
96 explicit DistributionMapping (const BoxArray& boxes,
97 int nprocs = ParallelDescriptor::NProcs());
98
99 explicit DistributionMapping (std::shared_ptr<Ref> a_ref);
100
106 const DistributionMapping& d2);
107
113 void define (const BoxArray& boxes, int nprocs = ParallelDescriptor::NProcs());
118 void define (const Vector<int>& pmap);
119 void define (Vector<int>&& pmap) noexcept;
126 [[nodiscard]] const Vector<int>& ProcessorMap () const noexcept;
127
129 [[nodiscard]] Long size () const noexcept { return Long(m_ref->m_pmap.size()); }
130 [[nodiscard]] Long capacity () const noexcept { return Long(m_ref->m_pmap.capacity()); }
131 [[nodiscard]] bool empty () const noexcept { return m_ref->m_pmap.empty(); }
132
134 [[nodiscard]] Long linkCount () const noexcept { return m_ref.use_count(); }
135
137 [[nodiscard]] int operator[] (int index) const noexcept { return m_ref->m_pmap[index]; }
138
139 std::istream& readFrom (std::istream& is);
140
141 std::ostream& writeOn (std::ostream& os) const;
142
144 static void strategy (Strategy how);
145
146 static Strategy strategy ();
147
149 static void SFC_Threshold (int n);
150
151 static int SFC_Threshold ();
152
154 bool operator== (const DistributionMapping& rhs) const noexcept;
155
157 bool operator!= (const DistributionMapping& rhs) const noexcept;
158
159 void SFCProcessorMap (const BoxArray& boxes, const std::vector<Long>& wgts, int nprocs,
160 bool sort=true);
161 void SFCProcessorMap (const BoxArray& boxes, const std::vector<Long>& wgts, int nprocs,
162 Real& efficiency, bool sort=true);
163 void KnapSackProcessorMap (const std::vector<Long>& wgts, int nprocs,
164 Real* efficiency=nullptr,
165 bool do_full_knapsack=true,
166 int nmax=std::numeric_limits<int>::max(),
167 bool sort=true);
168 void KnapSackProcessorMap (const DistributionMapping& olddm,
169 const std::vector<Long>& wgts, Real keep_ratio,
170 Real& old_efficiency, Real& new_efficiency,
171 int nmax=std::numeric_limits<int>::max());
172 void RoundRobinProcessorMap (int nboxes, int nprocs, bool sort=true);
173 void RoundRobinProcessorMap (const std::vector<Long>& wgts, int nprocs, bool sort=true);
174
185 static void Initialize ();
186
187 static void Finalize ();
188
189 static bool SameRefs (const DistributionMapping& lhs,
190 const DistributionMapping& rhs)
191 { return lhs.m_ref == rhs.m_ref; }
192
193 static DistributionMapping makeKnapSack (const MultiFab& weight,
194 int nmax=std::numeric_limits<int>::max());
195 static DistributionMapping makeKnapSack (const MultiFab& weight, Real& eff,
196 int nmax=std::numeric_limits<int>::max());
197 static DistributionMapping makeKnapSack (const Vector<Real>& rcost,
198 int nmax=std::numeric_limits<int>::max());
199 static DistributionMapping makeKnapSack (const Vector<Real>& rcost, Real& eff,
200 int nmax=std::numeric_limits<int>::max(),
201 bool sort=true);
202
227 static DistributionMapping makeKnapSack (const LayoutData<Real>& rcost_local,
228 Real& currentEfficiency, Real& proposedEfficiency,
229 int nmax=std::numeric_limits<int>::max(),
230 bool broadcastToAll=true,
232 Real keep_ratio = Real(0.0));
233
234 static DistributionMapping makeRoundRobin (const MultiFab& weight);
235 static DistributionMapping makeSFC (const MultiFab& weight, bool sort=true);
236 static DistributionMapping makeSFC (const MultiFab& weight, Real& eff, bool sort=true);
237 static DistributionMapping makeSFC (const Vector<Real>& rcost,
238 const BoxArray& ba, bool sort=true);
239 static DistributionMapping makeSFC (const Vector<Real>& rcost,
240 const BoxArray& ba, Real& eff, bool sort=true);
241
262 static DistributionMapping makeSFC (const LayoutData<Real>& rcost_local,
263 Real& currentEfficiency, Real& proposedEfficiency,
264 bool broadcastToAll=true,
266
271 static std::vector<std::vector<int> > makeSFC (const BoxArray& ba,
272 bool use_box_vol=true,
273 int nprocs=ParallelContext::NProcsSub() );
274
282 template <typename T>
284 const std::vector<T>& cost,
285 Real* efficiency);
286
289 static Vector<Long> ConvertCostRealToLong (const Vector<Real>& rcost);
290
291 [[nodiscard]] std::weak_ptr<Ref> getWeakRef () const;
292
293private:
294
295 const Vector<int>& getIndexArray ();
296 const std::vector<bool>& getOwnerShip ();
297
299 void RoundRobinProcessorMap (const BoxArray& boxes, int nprocs);
300 void KnapSackProcessorMap (const BoxArray& boxes, int nprocs);
301 void SFCProcessorMap (const BoxArray& boxes, int nprocs);
302 void RRSFCProcessorMap (const BoxArray& boxes, int nprocs);
303
304 using LIpair = std::pair<Long,int>;
305
306 struct LIpairLT
307 {
308 bool operator () (const LIpair& lhs,
309 const LIpair& rhs) const noexcept
310 {
311 return lhs.first < rhs.first;
312 }
313 };
314
315 struct LIpairGT
316 {
317 bool operator () (const LIpair& lhs,
318 const LIpair& rhs) const noexcept
319 {
320 return lhs.first > rhs.first;
321 }
322 };
323
324 static void Sort (std::vector<LIpair>& vec, bool reverse);
325
326 void RoundRobinDoIt (int nboxes,
327 int nprocs,
328 std::vector<LIpair>* LIpairV = nullptr,
329 bool sort = true);
330
331 void KnapSackDoIt (const std::vector<Long>& wgts,
332 int nprocs,
333 Real& efficiency,
334 bool do_full_knapsack,
335 int nmax=std::numeric_limits<int>::max(),
336 bool sort=true);
337
338 void SFCProcessorMapDoIt (const BoxArray& boxes,
339 const std::vector<Long>& wgts,
340 int nprocs,
341 bool sort=true,
342 Real* efficiency=nullptr);
343
344 void RRSFCDoIt (const BoxArray& boxes,
345 int nprocs);
346
348 static void LeastUsedCPUs (int nprocs, Vector<int>& result);
353 static void LeastUsedTeams (Vector<int>& rteam, Vector<Vector<int> >& rworker, int nteams, int nworkers);
354
356 using PVMF = void (DistributionMapping::*)(const BoxArray &, int);
357
359 static Strategy m_Strategy;
364 static PVMF m_BuildMap;
365
366 //
368 std::shared_ptr<Ref> m_ref;
369
370public:
371 struct RefID {
372 constexpr RefID () noexcept {} // =default does not work due to a clang bug // NOLINT
373 explicit RefID (Ref* data_) noexcept : data(data_) {}
374 bool operator< (const RefID& rhs) const noexcept { return std::less<>{}(data,rhs.data); }
375 bool operator== (const RefID& rhs) const noexcept { return data == rhs.data; }
376 bool operator!= (const RefID& rhs) const noexcept { return data != rhs.data; }
377 [[nodiscard]] const Ref *dataPtr() const noexcept { return data; }
378 void PrintPtr(std::ostream &os) const { os << data << '\n'; }
379 friend std::ostream& operator<< (std::ostream& os, const RefID& id);
380 private:
381 Ref* data = nullptr;
382 };
383
385 [[nodiscard]] RefID getRefID () const noexcept { return RefID { m_ref.get() }; }
386};
387
389
391std::ostream& operator<< (std::ostream& os, const DistributionMapping& pmap);
392
393std::ostream& operator<< (std::ostream& os, const DistributionMapping::RefID& id);
394
408DistributionMapping MakeSimilarDM (const BoxArray& ba, const MultiFab& mf, const IntVect& ng);
409
424DistributionMapping MakeSimilarDM (const BoxArray& ba, const BoxArray& src_ba,
425 const DistributionMapping& src_dm, const IntVect& ng);
426
427template <typename T>
429 const DistributionMapping& dm, const std::vector<T>& cost, Real* efficiency)
430{
431 const int nprocs = ParallelDescriptor::NProcs();
432 Vector<T> wgts(nprocs, T(0));
433
434 const auto nboxes = int(dm.size());
435 for (int ibox = 0; ibox < nboxes; ++ibox) {
436 wgts[dm[ibox]] += cost[ibox];
437 }
438
439 T max_weight = 0;
440 T sum_weight = 0;
441 for (auto const& w : wgts) {
442 max_weight = std::max(w, max_weight);
443 sum_weight += w;
444 }
445
446 AMREX_ASSERT(nprocs > 0 && max_weight > T(0));
447
448 *efficiency = static_cast<Real>(sum_weight) /
449 (static_cast<Real>(nprocs) * static_cast<Real>(max_weight));
450}
451
452}
453
454#endif /*BL_DISTRIBUTIONMAPPING_H*/
#define AMREX_ASSERT(EX)
Definition AMReX_BLassert.H:38
A collection of Boxes stored in an Array.
Definition AMReX_BoxArray.H:567
Calculates the distribution of FABs to MPI processes.
Definition AMReX_DistributionMapping.H:43
static void Initialize()
Initializes distribution strategy from ParmParse.
Definition AMReX_DistributionMapping.cpp:107
int operator[](int index) const noexcept
Equivalent to ProcessorMap()[index].
Definition AMReX_DistributionMapping.H:137
static DistributionMapping makeKnapSack(const MultiFab &weight, int nmax=std::numeric_limits< int >::max())
Definition AMReX_DistributionMapping.cpp:1722
Long linkCount() const noexcept
Number of references to this DistributionMapping.
Definition AMReX_DistributionMapping.H:134
Long size() const noexcept
Length of the underlying processor map.
Definition AMReX_DistributionMapping.H:129
void define(const BoxArray &boxes, int nprocs=ParallelDescriptor::NProcs())
Build mapping out of BoxArray over nprocs processors. You need to call this if you built your Distrib...
Definition AMReX_DistributionMapping.cpp:345
static bool SameRefs(const DistributionMapping &lhs, const DistributionMapping &rhs)
Definition AMReX_DistributionMapping.H:189
static int SFC_Threshold()
Definition AMReX_DistributionMapping.cpp:89
bool empty() const noexcept
Definition AMReX_DistributionMapping.H:131
bool operator!=(const DistributionMapping &rhs) const noexcept
Are the distributions different?
Definition AMReX_DistributionMapping.cpp:101
std::ostream & writeOn(std::ostream &os) const
Definition AMReX_DistributionMapping.cpp:1994
DistributionMapping() noexcept
The default constructor.
Definition AMReX_DistributionMapping.cpp:303
static Vector< Long > ConvertCostRealToLong(const Vector< Real > &rcost)
Definition AMReX_DistributionMapping.cpp:1589
static DistributionMapping makeRoundRobin(const MultiFab &weight)
Definition AMReX_DistributionMapping.cpp:1745
void KnapSackProcessorMap(const std::vector< Long > &wgts, int nprocs, Real *efficiency=nullptr, bool do_full_knapsack=true, int nmax=std::numeric_limits< int >::max(), bool sort=true)
Definition AMReX_DistributionMapping.cpp:855
bool operator==(const DistributionMapping &rhs) const noexcept
Are the distributions equal?
Definition AMReX_DistributionMapping.cpp:95
static DistributionMapping makeSFC(const MultiFab &weight, bool sort=true)
Definition AMReX_DistributionMapping.cpp:1756
void RoundRobinProcessorMap(int nboxes, int nprocs, bool sort=true)
Definition AMReX_DistributionMapping.cpp:455
void SFCProcessorMap(const BoxArray &boxes, const std::vector< Long > &wgts, int nprocs, bool sort=true)
Definition AMReX_DistributionMapping.cpp:1499
const Vector< int > & ProcessorMap() const noexcept
Returns a constant reference to the mapping of boxes in the underlying BoxArray to the CPU that holds...
Definition AMReX_DistributionMapping.cpp:47
RefID getRefID() const noexcept
This gives a unique ID of the reference, which is different from dmID above.
Definition AMReX_DistributionMapping.H:385
static Strategy strategy()
Definition AMReX_DistributionMapping.cpp:53
static void Finalize()
Definition AMReX_DistributionMapping.cpp:167
std::weak_ptr< Ref > getWeakRef() const
Definition AMReX_DistributionMapping.cpp:1940
Strategy
The distribution strategies.
Definition AMReX_DistributionMapping.H:50
@ UNDEFINED
Definition AMReX_DistributionMapping.H:50
@ KNAPSACK
Definition AMReX_DistributionMapping.H:50
@ RRSFC
Definition AMReX_DistributionMapping.H:50
@ ROUNDROBIN
Definition AMReX_DistributionMapping.H:50
@ SFC
Definition AMReX_DistributionMapping.H:50
Long capacity() const noexcept
Definition AMReX_DistributionMapping.H:130
std::istream & readFrom(std::istream &is)
Definition AMReX_DistributionMapping.cpp:1973
static void ComputeDistributionMappingEfficiency(const DistributionMapping &dm, const std::vector< T > &cost, Real *efficiency)
Computes the average cost per MPI rank given a distribution mapping global cost vector.
Definition AMReX_DistributionMapping.H:428
Base class for FabArray.
Definition AMReX_FabArrayBase.H:42
An Array of FortranArrayBox(FAB)-like Objects.
Definition AMReX_FabArray.H:347
a one-thingy-per-box distributed object
Definition AMReX_LayoutData.H:13
A collection (stored as an array) of FArrayBox objects.
Definition AMReX_MultiFab.H:40
This class is a thin wrapper around std::vector. Unlike vector, Vector::operator[] provides bound che...
Definition AMReX_Vector.H:28
amrex_real Real
Floating Point Type for Fields.
Definition AMReX_REAL.H:79
amrex_long Long
Definition AMReX_INT.H:30
int NProcs() noexcept
Definition AMReX_ParallelDescriptor.H:255
int IOProcessorNumber() noexcept
The MPI rank number of the I/O Processor (probably rank 0). This rank is usually used to write to std...
Definition AMReX_ParallelDescriptor.H:279
int NProcsSub() noexcept
number of ranks in current frame
Definition AMReX_ParallelContext.H:74
Definition AMReX_Amr.cpp:49
DistributionMapping MakeSimilarDM(const BoxArray &ba, const MultiFab &mf, const IntVect &ng)
Function that creates a DistributionMapping "similar" to that of a MultiFab.
Definition AMReX_DistributionMapping.cpp:2007
const int[]
Definition AMReX_BLProfiler.cpp:1664
std::ostream & operator<<(std::ostream &os, AmrMesh const &amr_mesh)
Definition AMReX_AmrMesh.cpp:1237
Definition AMReX_DistributionMapping.H:371
constexpr RefID() noexcept
Definition AMReX_DistributionMapping.H:372
friend std::ostream & operator<<(std::ostream &os, const RefID &id)
Definition AMReX_DistributionMapping.cpp:1966
bool operator==(const RefID &rhs) const noexcept
Definition AMReX_DistributionMapping.H:375
void PrintPtr(std::ostream &os) const
Definition AMReX_DistributionMapping.H:378
bool operator!=(const RefID &rhs) const noexcept
Definition AMReX_DistributionMapping.H:376
const Ref * dataPtr() const noexcept
Definition AMReX_DistributionMapping.H:377
bool operator<(const RefID &rhs) const noexcept
Definition AMReX_DistributionMapping.H:374
RefID(Ref *data_) noexcept
Definition AMReX_DistributionMapping.H:373
Definition AMReX_DistributionMapping.H:53
Ref()=default
Constructors to match those in DistributionMapping ....
std::vector< bool > m_ownership
true ownership
Definition AMReX_DistributionMapping.H:69
Ref(const Vector< int > &pmap)
Definition AMReX_DistributionMapping.H:59
Vector< int > m_pmap
index array for all boxes
Definition AMReX_DistributionMapping.H:67
Vector< int > m_index_array
index array for local boxes owned by the team
Definition AMReX_DistributionMapping.H:68
Ref(int len)
Definition AMReX_DistributionMapping.H:57
void clear()
dtor, copy-ctor, copy-op=, move-ctor, and move-op= are compiler generated.
Definition AMReX_DistributionMapping.H:65
Ref(Vector< int > &&pmap) noexcept
Definition AMReX_DistributionMapping.H:61