From eb4345ed44c5481f95fe494ab0dea9ed28822127 Mon Sep 17 00:00:00 2001
From: Mark Olesen <Mark.Olesen@esi-group.com>
Date: Fri, 14 Feb 2025 15:49:05 +0100
Subject: [PATCH] ENH: misc Pstream adjustments

- additional startup guard for inter-node/local-node queries (UPstream)

- impose linear communication tree for inter-node/local-node
  communicators. Was previously defaulted to a basic tree, but more
  consistent to have flat addressing for these types of connections.

- demand-driven UPstream::interNode_offsets() for walking
  inter-node ranges instead of creating it manually in various places.

- (style): List<int> instead of labelList for internal commsStruct
  since the communication structures are tied to MPI sizes
  and not to the OpenFOAM label sizes

- reduce the number of intermediate buffer allocations within
  gatherList, scatterList.
---
 .../test/nodeTopology/Test-nodeTopology.cxx   | 113 +++-----
 .../IOstreams/Pstreams/PstreamCombineGather.C |   6 +-
 .../db/IOstreams/Pstreams/PstreamGather.C     |  18 +-
 .../db/IOstreams/Pstreams/PstreamGatherList.C | 213 ++++++++++-----
 src/OpenFOAM/db/IOstreams/Pstreams/UPstream.C |  48 +++-
 src/OpenFOAM/db/IOstreams/Pstreams/UPstream.H |  77 +++---
 .../IOstreams/Pstreams/UPstreamCommsStruct.C  | 247 ++++++++++++------
 src/OpenFOAM/global/argList/argList.C         |  42 +--
 8 files changed, 479 insertions(+), 285 deletions(-)

diff --git a/applications/test/nodeTopology/Test-nodeTopology.cxx b/applications/test/nodeTopology/Test-nodeTopology.cxx
index db4a5eeaf4..9909d51d4f 100644
--- a/applications/test/nodeTopology/Test-nodeTopology.cxx
+++ b/applications/test/nodeTopology/Test-nodeTopology.cxx
@@ -59,13 +59,13 @@ int main(int argc, char *argv[])
 
     label nProcs = UPstream::nProcs(UPstream::worldComm);
 
-    List<int> interNodeProcs_fake;
+    DynamicList<int> fake_interNode_offsets;
 
     if (UPstream::parRun())
     {
         if (args.found("numProcs"))
         {
-            InfoErr<< "ignoring -np option in parallel" << nl;
+            InfoErr<< "ignoring -numProcs option in parallel" << nl;
         }
         if (args.found("cores"))
         {
@@ -78,25 +78,40 @@ int main(int argc, char *argv[])
         nProcs = args.getOrDefault<label>("numProcs", 16);
         label nCores = args.getOrDefault<label>("cores", 4);
 
+        auto& interNode_offsets = fake_interNode_offsets;
+
         if (nCores > 1 && nCores < nProcs)
         {
-            const label numNodes
-                = (nProcs/nCores) + ((nProcs % nCores) ? 1 : 0);
+            // Build the inter-node offsets
+            interNode_offsets.reserve((nProcs/nCores) + 4);
+            interNode_offsets.push_back(0);
 
-            interNodeProcs_fake.resize(numNodes);
-
-            for (label nodei = 0; nodei < numNodes; ++nodei)
+            for
+            (
+                int count = interNode_offsets.back() + nCores;
+                count < nProcs;
+                count += nCores
+            )
             {
-                interNodeProcs_fake[nodei] = nodei * nCores;
+                interNode_offsets.push_back(count);
             }
+
+            interNode_offsets.push_back(nProcs);
+        }
+        else
+        {
+            // Some fallback
+            interNode_offsets.reserve(2);
+            interNode_offsets.push_back(0);
+            interNode_offsets.push_back(nProcs);
         }
     }
 
-    const List<int>& interNodeProcs =
+    const List<int>& interNodeOffsets =
     (
         UPstream::parRun()
-      ? UPstream::procID(UPstream::commInterNode())
-      : interNodeProcs_fake
+      ? UPstream::interNode_offsets()
+      : fake_interNode_offsets
     );
 
 
@@ -111,79 +126,31 @@ int main(int argc, char *argv[])
         // Prefer left-to-right layout for large graphs
         os << indent << "rankdir=LR" << nl;
 
-        int pos = 0;
+        const label numNodes = interNodeOffsets.size()-1;
 
         // First level are the inter-node connections
-        const label parent = 0;
-        for (const auto proci : interNodeProcs)
         {
-            if (parent == proci) continue;
+            os  << indent << 0 << " -- " << token::LBRACE;
 
-            if (pos)
+            for (label nodei = 1; nodei < numNodes; ++nodei)
             {
-                os << "  ";
+                os  << ' ' << interNodeOffsets[nodei];
             }
-            else
-            {
-                os << indent;
-            }
-            os << parent << " -- " << proci;
 
-            if (++pos >= 4)  // Max 4 items per line
-            {
-                pos = 0;
-                os << nl;
-            }
+            os  << token::SPACE << token::RBRACE
+                << "  // inter-node: " << flatOutput(interNodeOffsets)
+                << nl;
         }
 
-        if (pos)
+        // Next level are the local-node connections
+        for (label nodei = 0; nodei < numNodes; ++nodei)
         {
-            pos = 0;
-            os << nl;
-        }
+            const auto firstProc = interNodeOffsets[nodei];
+            const auto lastProc = interNodeOffsets[nodei+1];
 
-        // Next level are within the nodes
-        for (label nodei = 0; nodei < interNodeProcs.size(); ++nodei)
-        {
-            pos = 0;
-
-            label firstProc = interNodeProcs[nodei];
-            const label lastProc =
-            (
-                (nodei+1 < interNodeProcs.size())
-              ? interNodeProcs[nodei+1]
-              : nProcs
-            );
-
-            os << indent << "// inter-node " << nodei
-                << " [" << firstProc
-                << ".." << lastProc-1 << "]" << nl;
-
-            for (label proci = firstProc; proci < lastProc; ++proci)
-            {
-                if (firstProc == proci) continue;
-
-                if (pos)
-                {
-                    os << "  ";
-                }
-                else
-                {
-                    os << indent;
-                }
-                os << firstProc << " -- " << proci;
-
-                if (++pos >= 4)  // Max 4 items per line
-                {
-                    pos = 0;
-                    os << nl;
-                }
-            }
-            if (pos)
-            {
-                pos = 0;
-                os << nl;
-            }
+            os  << indent << firstProc << " -- " << token::DQUOTE
+                << (firstProc+1) << ".." << (lastProc-1)
+                << token::DQUOTE << nl;
         }
 
         os.endBlock();
diff --git a/src/OpenFOAM/db/IOstreams/Pstreams/PstreamCombineGather.C b/src/OpenFOAM/db/IOstreams/Pstreams/PstreamCombineGather.C
index 911bc8ebb8..f91a9ffae5 100644
--- a/src/OpenFOAM/db/IOstreams/Pstreams/PstreamCombineGather.C
+++ b/src/OpenFOAM/db/IOstreams/Pstreams/PstreamCombineGather.C
@@ -60,7 +60,7 @@ void Foam::Pstream::combineGather
         const auto& myComm = comms[UPstream::myProcNo(comm)];
 
         // Receive from my downstairs neighbours
-        for (const label belowID : myComm.below())
+        for (const auto belowID : myComm.below())
         {
             if constexpr (is_contiguous_v<T>)
             {
@@ -172,7 +172,7 @@ void Foam::Pstream::listCombineGather
         const auto& myComm = comms[UPstream::myProcNo(comm)];
 
         // Receive from my downstairs neighbours
-        for (const label belowID : myComm.below())
+        for (const auto belowID : myComm.below())
         {
             if constexpr (is_contiguous_v<T>)
             {
@@ -288,7 +288,7 @@ void Foam::Pstream::mapCombineGather
         const auto& myComm = comms[UPstream::myProcNo(comm)];
 
         // Receive from my downstairs neighbours
-        for (const label belowID : myComm.below())
+        for (const auto belowID : myComm.below())
         {
             // Map/HashTable: non-contiguous
 
diff --git a/src/OpenFOAM/db/IOstreams/Pstreams/PstreamGather.C b/src/OpenFOAM/db/IOstreams/Pstreams/PstreamGather.C
index f3ec676e1e..e28bc0bb4c 100644
--- a/src/OpenFOAM/db/IOstreams/Pstreams/PstreamGather.C
+++ b/src/OpenFOAM/db/IOstreams/Pstreams/PstreamGather.C
@@ -44,18 +44,18 @@ void Foam::Pstream::gather
     T& value,
     const BinaryOp& bop,
     const int tag,
-    const label comm
+    const label communicator
 )
 {
-    if (UPstream::is_parallel(comm))
+    if (UPstream::is_parallel(communicator))
     {
         // Communication order
-        const auto& comms = UPstream::whichCommunication(comm);
+        const auto& comms = UPstream::whichCommunication(communicator);
         // if (comms.empty()) return;  // extra safety?
-        const auto& myComm = comms[UPstream::myProcNo(comm)];
+        const auto& myComm = comms[UPstream::myProcNo(communicator)];
 
         // Receive from my downstairs neighbours
-        for (const label belowID : myComm.below())
+        for (const auto belowID : myComm.below())
         {
             T received;
 
@@ -68,12 +68,12 @@ void Foam::Pstream::gather
                     reinterpret_cast<char*>(&received),
                     sizeof(T),
                     tag,
-                    comm
+                    communicator
                 );
             }
             else
             {
-                IPstream::recv(received, belowID, tag, comm);
+                IPstream::recv(received, belowID, tag, communicator);
             }
 
             value = bop(value, received);
@@ -91,12 +91,12 @@ void Foam::Pstream::gather
                     reinterpret_cast<const char*>(&value),
                     sizeof(T),
                     tag,
-                    comm
+                    communicator
                 );
             }
             else
             {
-                OPstream::send(value, myComm.above(), tag, comm);
+                OPstream::send(value, myComm.above(), tag, communicator);
             }
         }
     }
diff --git a/src/OpenFOAM/db/IOstreams/Pstreams/PstreamGatherList.C b/src/OpenFOAM/db/IOstreams/Pstreams/PstreamGatherList.C
index edc269a067..67e2cafa4e 100644
--- a/src/OpenFOAM/db/IOstreams/Pstreams/PstreamGatherList.C
+++ b/src/OpenFOAM/db/IOstreams/Pstreams/PstreamGatherList.C
@@ -32,13 +32,13 @@ Description
     values[UPstream::myProcNo(comm)].
     Note: after gather every processor only knows its own data and that of the
     processors below it. Only the 'master' of the communication schedule holds
-    a fully filled List. Use scatter to distribute the data.
+    a fully filled List. Use broadcast to distribute the data.
 
 \*---------------------------------------------------------------------------*/
 
+#include "contiguous.H"
 #include "IPstream.H"
 #include "OPstream.H"
-#include "contiguous.H"
 
 // * * * * * * * * * * * * * * * Member Functions  * * * * * * * * * * * * * //
 
@@ -48,13 +48,13 @@ void Foam::Pstream::gatherList
     const UPstream::commsStructList& comms,
     UList<T>& values,
     const int tag,
-    const label comm
+    const label communicator
 )
 {
-    if (!comms.empty() && UPstream::is_parallel(comm))
+    if (!comms.empty() && UPstream::is_parallel(communicator))
     {
-        const label myProci = UPstream::myProcNo(comm);
-        const label numProc = UPstream::nProcs(comm);
+        const label myProci = UPstream::myProcNo(communicator);
+        const label numProc = UPstream::nProcs(communicator);
 
         if (values.size() < numProc)
         {
@@ -67,29 +67,71 @@ void Foam::Pstream::gatherList
         // My communication order
         const auto& myComm = comms[myProci];
 
-        // Receive from my downstairs neighbours
-        for (const label belowID : myComm.below())
+
+        // Local buffer for send/recv of contiguous
+        [[maybe_unused]] DynamicList<T> buffer;
+
+        // Presize buffer
+        if constexpr (is_contiguous_v<T>)
         {
-            const labelList& belowLeaves = comms[belowID].allBelow();
+            label maxCount = 0;
+
+            for (const auto belowID : myComm.below())
+            {
+                auto count = comms[belowID].allBelow().size();
+                maxCount = Foam::max(maxCount, count);
+            }
+
+            if (myComm.above() >= 0)
+            {
+                auto count = myComm.allBelow().size();
+                maxCount = Foam::max(maxCount, count);
+            }
+
+            buffer.reserve(maxCount + 1);
+        }
+
+
+        // Receive from my downstairs neighbours
+        for (const auto belowID : myComm.below())
+        {
+            const auto& leaves = comms[belowID].allBelow();
 
             if constexpr (is_contiguous_v<T>)
             {
-                List<T> received(belowLeaves.size() + 1);
-
-                UIPstream::read
-                (
-                    UPstream::commsTypes::scheduled,
-                    belowID,
-                    received,
-                    tag,
-                    comm
-                );
-
-                values[belowID] = received[0];
-
-                forAll(belowLeaves, leafI)
+                if (leaves.empty())
                 {
-                    values[belowLeaves[leafI]] = received[leafI + 1];
+                    // Receive directly into destination
+                    UIPstream::read
+                    (
+                        UPstream::commsTypes::scheduled,
+                        belowID,
+                        values[belowID],
+                        tag,
+                        communicator
+                    );
+                }
+                else
+                {
+                    // Receive via intermediate buffer
+                    buffer.resize_nocopy(leaves.size() + 1);
+
+                    UIPstream::read
+                    (
+                        UPstream::commsTypes::scheduled,
+                        belowID,
+                        buffer,
+                        tag,
+                        communicator
+                    );
+
+                    label recvIdx(0);
+                    values[belowID] = buffer[recvIdx++];
+
+                    for (const auto leafID : leaves)
+                    {
+                        values[leafID] = buffer[recvIdx++];
+                    }
                 }
             }
             else
@@ -100,7 +142,7 @@ void Foam::Pstream::gatherList
                     belowID,
                     0,  // bufsize
                     tag,
-                    comm
+                    communicator
                 );
                 fromBelow >> values[belowID];
 
@@ -112,7 +154,7 @@ void Foam::Pstream::gatherList
                 }
 
                 // Receive from all other processors below belowID
-                for (const label leafID : belowLeaves)
+                for (const auto leafID : leaves)
                 {
                     fromBelow >> values[leafID];
 
@@ -131,7 +173,7 @@ void Foam::Pstream::gatherList
         // - all belowLeaves next
         if (myComm.above() >= 0)
         {
-            const labelList& belowLeaves = myComm.allBelow();
+            const auto& leaves = myComm.allBelow();
 
             if (debug & 2)
             {
@@ -142,22 +184,40 @@ void Foam::Pstream::gatherList
 
             if constexpr (is_contiguous_v<T>)
             {
-                List<T> sending(belowLeaves.size() + 1);
-                sending[0] = values[myProci];
-
-                forAll(belowLeaves, leafI)
+                if (leaves.empty())
                 {
-                    sending[leafI + 1] = values[belowLeaves[leafI]];
+                    // Send directly
+                    UOPstream::write
+                    (
+                        UPstream::commsTypes::scheduled,
+                        myComm.above(),
+                        values[myProci],
+                        tag,
+                        communicator
+                    );
                 }
+                else
+                {
+                    // Send via intermediate buffer
+                    buffer.resize_nocopy(leaves.size() + 1);
 
-                UOPstream::write
-                (
-                    UPstream::commsTypes::scheduled,
-                    myComm.above(),
-                    sending,
-                    tag,
-                    comm
-                );
+                    label sendIdx(0);
+                    buffer[sendIdx++] = values[myProci];
+
+                    for (const auto leafID : leaves)
+                    {
+                        buffer[sendIdx++] = values[leafID];
+                    }
+
+                    UOPstream::write
+                    (
+                        UPstream::commsTypes::scheduled,
+                        myComm.above(),
+                        buffer,
+                        tag,
+                        communicator
+                    );
+                }
             }
             else
             {
@@ -167,11 +227,11 @@ void Foam::Pstream::gatherList
                     myComm.above(),
                     0,  // bufsize
                     tag,
-                    comm
+                    communicator
                 );
                 toAbove << values[myProci];
 
-                for (const label leafID : belowLeaves)
+                for (const auto leafID : leaves)
                 {
                     if (debug & 2)
                     {
@@ -193,17 +253,17 @@ void Foam::Pstream::scatterList
     const UPstream::commsStructList& comms,
     UList<T>& values,
     const int tag,
-    const label comm
+    const label communicator
 )
 {
     // Apart from the additional size check, the only difference
     // between scatterList() and using broadcast(List<T>&) or a regular
     // scatter(List<T>&) is that processor-local data is skipped.
 
-    if (!comms.empty() && UPstream::is_parallel(comm))
+    if (!comms.empty() && UPstream::is_parallel(communicator))
     {
-        const label myProci = UPstream::myProcNo(comm);
-        const label numProc = UPstream::nProcs(comm);
+        const label myProci = UPstream::myProcNo(communicator);
+        const label numProc = UPstream::nProcs(communicator);
 
         if (values.size() < numProc)
         {
@@ -216,27 +276,53 @@ void Foam::Pstream::scatterList
         // My communication order
         const auto& myComm = comms[myProci];
 
+
+        // Local buffer for send/recv of contiguous
+        [[maybe_unused]] DynamicList<T> buffer;
+
+        // Presize buffer
+        if constexpr (is_contiguous_v<T>)
+        {
+            label maxCount = 0;
+
+            if (myComm.above() >= 0)
+            {
+                auto count = myComm.allNotBelow().size();
+                maxCount = Foam::max(maxCount, count);
+            }
+
+            for (const auto belowID : myComm.below())
+            {
+                auto count = comms[belowID].allNotBelow().size();
+                maxCount = Foam::max(maxCount, count);
+            }
+
+            buffer.reserve(maxCount);
+        }
+
+
         // Receive from up
         if (myComm.above() >= 0)
         {
-            const labelList& notBelowLeaves = myComm.allNotBelow();
+            const auto& leaves = myComm.allNotBelow();
 
             if constexpr (is_contiguous_v<T>)
             {
-                List<T> received(notBelowLeaves.size());
+                buffer.resize_nocopy(leaves.size());
 
                 UIPstream::read
                 (
                     UPstream::commsTypes::scheduled,
                     myComm.above(),
-                    received,
+                    buffer,
                     tag,
-                    comm
+                    communicator
                 );
 
-                forAll(notBelowLeaves, leafI)
+                label recvIdx(0);
+                for (const auto leafID : leaves)
                 {
-                    values[notBelowLeaves[leafI]] = received[leafI];
+                    values[leafID] = buffer[recvIdx++];
                 }
             }
             else
@@ -247,10 +333,10 @@ void Foam::Pstream::scatterList
                     myComm.above(),
                     0,  // bufsize
                     tag,
-                    comm
+                    communicator
                 );
 
-                for (const label leafID : notBelowLeaves)
+                for (const auto leafID : leaves)
                 {
                     fromAbove >> values[leafID];
 
@@ -267,25 +353,26 @@ void Foam::Pstream::scatterList
         // Send to my downstairs neighbours
         forAllReverse(myComm.below(), belowI)
         {
-            const label belowID = myComm.below()[belowI];
-            const labelList& notBelowLeaves = comms[belowID].allNotBelow();
+            const auto belowID = myComm.below()[belowI];
+            const auto& leaves = comms[belowID].allNotBelow();
 
             if constexpr (is_contiguous_v<T>)
             {
-                List<T> sending(notBelowLeaves.size());
+                buffer.resize_nocopy(leaves.size());
 
-                forAll(notBelowLeaves, leafI)
+                label sendIdx(0);
+                for (const auto leafID : leaves)
                 {
-                    sending[leafI] = values[notBelowLeaves[leafI]];
+                    buffer[sendIdx++] = values[leafID];
                 }
 
                 UOPstream::write
                 (
                     UPstream::commsTypes::scheduled,
                     belowID,
-                    sending,
+                    buffer,
                     tag,
-                    comm
+                    communicator
                 );
             }
             else
@@ -296,11 +383,11 @@ void Foam::Pstream::scatterList
                     belowID,
                     0,  // bufsize
                     tag,
-                    comm
+                    communicator
                 );
 
                 // Send data destined for all other processors below belowID
-                for (const label leafID : notBelowLeaves)
+                for (const auto leafID : leaves)
                 {
                     toBelow << values[leafID];
 
diff --git a/src/OpenFOAM/db/IOstreams/Pstreams/UPstream.C b/src/OpenFOAM/db/IOstreams/Pstreams/UPstream.C
index 4f47d5f983..cd2e1d3c84 100644
--- a/src/OpenFOAM/db/IOstreams/Pstreams/UPstream.C
+++ b/src/OpenFOAM/db/IOstreams/Pstreams/UPstream.C
@@ -640,7 +640,7 @@ Foam::UPstream::treeCommunication(const label communicator)
 }
 
 
-void Foam::UPstream::printCommTree(const label communicator)
+void Foam::UPstream::printCommTree(int communicator)
 {
     const auto& comms = UPstream::whichCommunication(communicator);
 
@@ -663,14 +663,60 @@ bool Foam::UPstream::usingNodeComms(const label communicator)
     (
         parRun_ && (constWorldComm_ == communicator)
      && (nodeCommsControl_ > 0)
+
         // More than one node and above defined threshold
      && (numNodes_ > 1) && (numNodes_ >= nodeCommsMin_)
         // Some processes do share nodes
      && (numNodes_ < procIDs_[constWorldComm_].size())
+
+        // Extra paranoid (guard against calling during startup)
+     && (commInterNode_ > constWorldComm_)
+     && (commLocalNode_ > constWorldComm_)
     );
 }
 
 
+const Foam::List<int>& Foam::UPstream::interNode_offsets()
+{
+    static std::unique_ptr<List<int>> singleton;
+
+    if (!singleton)
+    {
+        // Extra paranoid (guard against calling during startup)
+        if
+        (
+            (commInterNode_ <= constWorldComm_)
+         || (commInterNode_ >= procIDs_.size())
+        )
+        {
+            return List<int>::null();
+        }
+
+        singleton = std::make_unique<List<int>>();
+        auto& offsets = *singleton;
+
+        const auto& procs = procIDs_[commInterNode_];
+
+        // The procIDs_ are already the offsets, but missing the end offset
+        if (!procs.empty())
+        {
+            const auto count = procs.size();
+
+            offsets.resize(count+1);
+            std::copy_n
+            (
+                procs.begin(),
+                count,
+                offsets.begin()
+            );
+            offsets[count] = UPstream::nProcs(constWorldComm_);
+        }
+    }
+
+    return *singleton;
+}
+
+
 // * * * * * * * * * * * * * * Static Data Members * * * * * * * * * * * * * //
 
 bool Foam::UPstream::parRun_(false);
diff --git a/src/OpenFOAM/db/IOstreams/Pstreams/UPstream.H b/src/OpenFOAM/db/IOstreams/Pstreams/UPstream.H
index b7df5c279e..1030d307e3 100644
--- a/src/OpenFOAM/db/IOstreams/Pstreams/UPstream.H
+++ b/src/OpenFOAM/db/IOstreams/Pstreams/UPstream.H
@@ -108,18 +108,18 @@ public:
             // Private Data
 
                 //- The procID of the processor \em directly above
-                label above_;
+                int above_;
 
                 //- The procIDs of processors \em directly below
-                labelList below_;
+                List<int> below_;
 
                 //- The procIDs of all processors below myProcNo,
                 //- not just directly below
-                labelList allBelow_;
+                List<int> allBelow_;
 
                 //- The procIDs of all processors not below myProcNo
-                //  (inverse of allBelow_ without myProcNo)
-                labelList allNotBelow_;
+                //- (inverse of allBelow_ without myProcNo)
+                List<int> allNotBelow_;
 
 
         public:
@@ -132,20 +132,20 @@ public:
                 //- Move construct from components
                 commsStruct
                 (
-                    const label above,
-                    labelList&& below,
-                    labelList&& allBelow,
-                    labelList&& allNotBelow
+                    const int above,
+                    List<int>&& below,
+                    List<int>&& allBelow,
+                    List<int>&& allNotBelow
                 );
 
                 //- Copy construct from below, allBelow components
                 commsStruct
                 (
-                    const label numProcs,
-                    const label myProcID,
-                    const label above,
-                    const labelUList& below,
-                    const labelUList& allBelow
+                    const int numProcs,
+                    const int myProcID,
+                    const int above,
+                    const UList<int>& below,
+                    const UList<int>& allBelow
                 );
 
 
@@ -153,26 +153,26 @@ public:
 
             // Access
 
-                //- The number of processors addressed by the structure
-                label nProcs() const noexcept;
-
                 //- The procID of the processor \em directly above
-                label above() const noexcept { return above_; }
+                int above() const noexcept { return above_; }
 
                 //- The procIDs of the processors \em directly below
-                const labelList& below() const noexcept { return below_; }
+                const List<int>& below() const noexcept { return below_; }
 
-                //- The procIDs of all processors below
+                //- The procIDs of \em all processors below
                 //- (so not just directly below)
-                const labelList& allBelow() const noexcept { return allBelow_; }
+                const List<int>& allBelow() const noexcept { return allBelow_; }
 
                 //- The procIDs of all processors not below myProcNo.
                 //- The inverse set of allBelow without myProcNo.
-                const labelList& allNotBelow() const noexcept
+                const List<int>& allNotBelow() const noexcept
                 {
                     return allNotBelow_;
                 }
 
+                //- The number of processors addressed by the structure
+                int nProcs() const noexcept;
+
 
             // Edit
 
@@ -183,9 +183,9 @@ public:
                 //- possibly with communicator-specific adjustments
                 void reset
                 (
-                    const label procID,
-                    const label numProcs,
-                    const label comm = -1
+                    const int myProci,
+                    const int numProcs,
+                    const int communicator
                 );
 
 
@@ -203,7 +203,7 @@ public:
             // Private Data
 
                 //- The communicator index
-                label comm_;
+                int comm_;
 
                 //- The communication tree
                 List<commsStruct> tree_;
@@ -216,7 +216,7 @@ public:
                 commsStructList() noexcept : comm_(-1) {}
 
                 //- Construct empty with given communicator
-                commsStructList(label comm) noexcept : comm_(comm) {}
+                explicit commsStructList(int comm) noexcept : comm_(comm) {}
 
 
             // Static Functions
@@ -230,8 +230,8 @@ public:
                 //- True if communicator is non-negative (ie, was assigned)
                 bool good() const noexcept { return (comm_ >= 0); }
 
-                //- The communicator label
-                label comm() const noexcept { return comm_; }
+                //- The communicator internal index
+                int comm() const noexcept { return comm_; }
 
                 //- Clear the list
                 void clear() { return tree_.clear(); }
@@ -242,20 +242,23 @@ public:
                 //- The number of entries
                 label size() const noexcept { return tree_.size(); }
 
-                //- Reset communicator index and clear demand-driven entries
-                void init(const label comm);
+                //- Reset communicator index, fill tree with empty entries
+                void init(int communicator);
+
+                //- Reset communicator index, clear tree entries
+                void reset(int communicator);
 
                 //- Get existing or create (demand-driven) entry
-                const UPstream::commsStruct& get(const label proci) const;
+                const UPstream::commsStruct& get(int proci) const;
 
                 //- Get existing or create (demand-driven) entry
-                const UPstream::commsStruct& operator[](const label proci) const
+                const UPstream::commsStruct& operator[](int proci) const
                 {
                     return get(proci);
                 }
 
                 //- Print un-directed graph in graphviz dot format
-                void printGraph(Ostream& os, label proci = 0) const;
+                void printGraph(Ostream& os, int proci = 0) const;
         };
 
 
@@ -1074,6 +1077,10 @@ public:
             return rangeType(1, static_cast<int>(nProcs(communicator)-1));
         }
 
+        //- Processor offsets corresponding to the inter-node communicator
+        static const List<int>& interNode_offsets();
+
+
         //- Communication schedule for linear all-to-master (proc 0)
         static const commsStructList& linearCommunication
         (
@@ -1105,7 +1112,7 @@ public:
             (
                 np <= 1
               ? commsStructList::null()
-              : (np <= 2 || np < nProcsSimpleSum)
+              : (np <= 2 || np < UPstream::nProcsSimpleSum)
               ? linearCommunication(communicator)
               : treeCommunication(communicator)
             );
diff --git a/src/OpenFOAM/db/IOstreams/Pstreams/UPstreamCommsStruct.C b/src/OpenFOAM/db/IOstreams/Pstreams/UPstreamCommsStruct.C
index 72a41bba26..e55eff05ed 100644
--- a/src/OpenFOAM/db/IOstreams/Pstreams/UPstreamCommsStruct.C
+++ b/src/OpenFOAM/db/IOstreams/Pstreams/UPstreamCommsStruct.C
@@ -28,6 +28,9 @@ License
 
 #include "UPstream.H"
 
+#include <algorithm>
+#include <numeric>
+
 // * * * * * * * * * * * * * * * Local Functions * * * * * * * * * * * * * * //
 
 namespace Foam
@@ -38,9 +41,9 @@ static void printGraph_impl
 (
     Ostream& os,
     const UPstream::commsStructList& comms,
-    const label proci,
-    label depth,
-    const label maxDepth = 1024
+    const int proci,
+    int depth,
+    const int maxDepth = 1024
 )
 {
     if (proci >= comms.size())
@@ -59,41 +62,80 @@ static void printGraph_impl
 
         // Prefer left-to-right layout for large graphs
         os << indent << "rankdir=LR" << nl;
+    }
 
-        if (below.empty())
+
+    // Output the immediate neighbours below
+
+    if (below.empty())
+    {
+        if (proci == 0)
         {
             // A graph with a single-node (eg, self-comm)
             os << indent << proci << nl;
         }
     }
-
-    int pos = 0;
-
-    for (const auto nbrProci : below)
+    else
     {
-        if (pos)
-        {
-            os << "  ";
-        }
-        else
-        {
-            os << indent;
-        }
-        os << proci << " -- " << nbrProci;
+        os << indent << proci << " -- " << token::BEGIN_BLOCK;
 
-        if (++pos >= 4)  // Max 4 items per line
+        // Accumulate into ranges whenever possible
+        IntRange<int> range;
+
+        // Print accumulated range and reset
+        auto emit_range = [&]()
         {
-            pos = 0;
-            os << nl;
+            if (!range.empty())
+            {
+                os << ' ';
+                if (range.min() < range.max())
+                {
+                    os << '"' << range.min() << ".." << range.max() << '"';
+                }
+                else
+                {
+                    os << range.min();
+                }
+                range.reset();
+            }
+        };
+
+        for (const auto nbrProci : below)
+        {
+            const bool terminal = comms[nbrProci].below().empty();
+
+            if
+            (
+                terminal
+             && (!range.empty() && (range.max()+1 == nbrProci))
+            )
+            {
+                // Accumulate
+                ++range;
+                continue;
+            }
+
+            // Emit accumulated range
+            emit_range();
+
+            if (terminal)
+            {
+                range.reset(nbrProci, 1);
+            }
+            else
+            {
+                os << token::SPACE << nbrProci;
+            }
         }
+
+        // Emit accumulated range
+        emit_range();
+
+        os << token::SPACE << token::END_BLOCK << nl;
     }
 
-    if (pos)
-    {
-        os << nl;
-    }
 
-    // Limit the maximum depth
+    // Recurse into below neighbours, but limit the maximum depth
     ++depth;
     if (depth >= maxDepth && (proci != 0))
     {
@@ -109,7 +151,6 @@ static void printGraph_impl
     if (proci == 0)
     {
         os.endBlock();
-
         os << "// end graph" << nl;
     }
 }
@@ -150,46 +191,46 @@ static void printGraph_impl
 namespace Foam
 {
 
-static label simpleTree
+static int simpleTree
 (
-    const label procID,
-    const label numProcs,
+    const int myProci,
+    const int numProcs,
 
-    DynamicList<label>& below,
-    DynamicList<label>& allBelow
+    DynamicList<int>& below,
+    DynamicList<int>& allBelow
 )
 {
-    label above(-1);
+    int above(-1);
 
-    for (label mod = 2, step = 1; step < numProcs; step = mod)
+    for (int mod = 2, step = 1; step < numProcs; step = mod)
     {
         mod = step * 2;
 
-        if (procID % mod)
+        if (myProci % mod)
         {
             // The rank above
-            above = procID - (procID % mod);
+            above = myProci - (myProci % mod);
             break;
         }
         else
         {
             for
             (
-                label j = procID + step;
-                j < numProcs && j < procID + mod;
-                j += step
+                int i = myProci + step;
+                i < numProcs && i < myProci + mod;
+                i += step
             )
             {
-                below.push_back(j);
+                below.push_back(i);
             }
             for
             (
-                label j = procID + step;
-                j < numProcs && j < procID + mod;
-                j++
+                int i = myProci + step;
+                i < numProcs && i < myProci + mod;
+                ++i
             )
             {
-                allBelow.push_back(j);
+                allBelow.push_back(i);
             }
         }
     }
@@ -204,10 +245,10 @@ static label simpleTree
 
 Foam::UPstream::commsStruct::commsStruct
 (
-    const label above,
-    labelList&& below,
-    labelList&& allBelow,
-    labelList&& allNotBelow
+    const int above,
+    List<int>&& below,
+    List<int>&& allBelow,
+    List<int>&& allNotBelow
 )
 :
     above_(above),
@@ -219,11 +260,11 @@ Foam::UPstream::commsStruct::commsStruct
 
 Foam::UPstream::commsStruct::commsStruct
 (
-    const label numProcs,
-    const label myProcID,
-    const label above,
-    const labelUList& below,
-    const labelUList& allBelow
+    const int numProcs,
+    const int myProcID,
+    const int above,
+    const UList<int>& below,
+    const UList<int>& allBelow
 )
 :
     above_(above),
@@ -237,14 +278,14 @@ Foam::UPstream::commsStruct::commsStruct
     isNotBelow[myProcID] = false;
 
     // Exclude allBelow
-    for (const label proci : allBelow)
+    for (const auto proci : allBelow)
     {
         isNotBelow[proci] = false;
     }
 
     // Compacting to obtain allNotBelow_
-    label nNotBelow = 0;
-    forAll(isNotBelow, proci)
+    int nNotBelow = 0;
+    for (int proci = 0; proci < numProcs; ++proci)
     {
         if (isNotBelow[proci])
         {
@@ -266,7 +307,7 @@ Foam::UPstream::commsStruct::commsStruct
 void Foam::UPstream::commsStructList::printGraph
 (
     Ostream& os,
-    const label proci
+    const int proci
 ) const
 {
     // Print graph - starting at depth 0
@@ -282,9 +323,9 @@ void Foam::UPstream::commsStructList::printGraph
 
 // * * * * * * * * * * * * * * * Member Functions  * * * * * * * * * * * * * //
 
-Foam::label Foam::UPstream::commsStruct::nProcs() const noexcept
+int Foam::UPstream::commsStruct::nProcs() const noexcept
 {
-    return (1 + allBelow_.size() + allNotBelow_.size());
+    return (1 + int(allBelow_.size() + allNotBelow_.size()));
 }
 
 
@@ -299,46 +340,65 @@ void Foam::UPstream::commsStruct::reset()
 
 void Foam::UPstream::commsStruct::reset
 (
-    const label procID,
-    const label numProcs,
-    [[maybe_unused]] const label comm
+    const int myProci,
+    const int numProcs,
+    const int communicator
 )
 {
     reset();
 
-    if (numProcs <= 2 || numProcs < UPstream::nProcsSimpleSum)
+    // Linear (flat) communication pattern
+    if
+    (
+        // Trivially small domains
+        (numProcs <= 2 || numProcs < UPstream::nProcsSimpleSum)
+
+        // local-node: assume that the local communication is low-latency
+     || (
+            UPstream::commLocalNode() == communicator
+         && UPstream::commLocalNode() > UPstream::commConstWorld()
+        )
+        // inter-node: presumably relatively few nodes and/or
+        //     higher latency with larger messages being sent
+     || (
+            UPstream::commInterNode() == communicator
+         && UPstream::commInterNode() > UPstream::commConstWorld()
+        )
+    )
     {
         // Linear communication pattern
-        label above(-1);
-        labelList below;
+        int above(-1);
+        List<int> below;
 
-        if (procID == 0)
+        if (myProci == 0)
         {
-            below = identity(numProcs-1, 1);
+            below.resize(numProcs-1);
+            std::iota(below.begin(), below.end(), 1);
         }
         else
         {
             above = 0;
         }
 
-        *this = UPstream::commsStruct(numProcs, procID, above, below, below);
+        *this = UPstream::commsStruct(numProcs, myProci, above, below, below);
         return;
     }
 
 
-    // Simple tree communication pattern
-    DynamicList<label> below;
-    DynamicList<label> allBelow;
 
-    label above = simpleTree
+    DynamicList<int> below;
+    DynamicList<int> allBelow;
+
+    // Simple tree communication pattern
+    int above = simpleTree
     (
-        procID,
+        myProci,
         numProcs,
         below,
         allBelow
     );
 
-    *this = UPstream::commsStruct(numProcs, procID, above, below, allBelow);
+    *this = UPstream::commsStruct(numProcs, myProci, above, below, allBelow);
 }
 
 
@@ -360,19 +420,36 @@ Foam::UPstream::commsStructList::null()
 
 // * * * * * * * * * * * * * * * Member Functions  * * * * * * * * * * * * * //
 
-void Foam::UPstream::commsStructList::init(const label comm)
+void Foam::UPstream::commsStructList::init(int communicator)
 {
-    comm_ = comm;
+    comm_ = communicator;
+    tree_.clear();
+    if (comm_ >= 0)
+    {
+        tree_.resize(UPstream::nProcs(comm_));
+    }
+}
+
+
+void Foam::UPstream::commsStructList::reset(int communicator)
+{
+    comm_ = communicator;
     tree_.clear();
-    tree_.resize(UPstream::nProcs(comm));
 }
 
 
 const Foam::UPstream::commsStruct&
-Foam::UPstream::commsStructList::get(const label proci) const
+Foam::UPstream::commsStructList::get(int proci) const
 {
+    const auto numProcs = UPstream::nProcs(comm_);
+
+    // Only if reset(comm) instead of init(comm) was used
+    if (tree_.size() < numProcs)
+    {
+        const_cast<List<commsStruct>&>(tree_).resize(numProcs);
+    }
+
     const UPstream::commsStruct& entry = tree_[proci];
-    const auto numProcs = tree_.size();
 
     if (entry.nProcs() != numProcs)
     {
@@ -391,10 +468,8 @@ bool Foam::UPstream::commsStruct::operator==(const commsStruct& comm) const
 {
     return
     (
-        (above_ == comm.above())
-     && (below_ == comm.below())
-     // && (allBelow_ == comm.allBelow())
-     // && (allNotBelow_ == comm.allNotBelow())
+        (above() == comm.above())
+     && (below() == comm.below())
     );
 }
 
@@ -409,10 +484,10 @@ bool Foam::UPstream::commsStruct::operator!=(const commsStruct& comm) const
 
 Foam::Ostream& Foam::operator<<(Ostream& os, const UPstream::commsStruct& comm)
 {
-    os  << comm.above() << nl << token::SPACE << token::SPACE;
-    comm.below().writeList(os) << nl << token::SPACE << token::SPACE;
-    comm.allBelow().writeList(os) << nl << token::SPACE << token::SPACE;
-    comm.allNotBelow().writeList(os);
+    os  << comm.above() << nl;
+    os  << "  "; comm.below().writeList(os) << nl;
+    os  << "  "; comm.allBelow().writeList(os) << nl;
+    os  << "  "; comm.allNotBelow().writeList(os);
 
     os.check(FUNCTION_NAME);
     return os;
diff --git a/src/OpenFOAM/global/argList/argList.C b/src/OpenFOAM/global/argList/argList.C
index 8abb9a5457..99bee6f6fd 100644
--- a/src/OpenFOAM/global/argList/argList.C
+++ b/src/OpenFOAM/global/argList/argList.C
@@ -2110,21 +2110,33 @@ void Foam::argList::parse
             Info<< " (" << UPstream::nProcs() << " ranks, "
                 << UPstream::numNodes() << " nodes)" << nl;
 
-            Info<< "    floatTransfer      : "
-                << Switch::name(UPstream::floatTransfer) << nl
-                << "    maxCommsSize       : "
-                << UPstream::maxCommsSize << nl
-                << "    nProcsSimpleSum    : "
-                << UPstream::nProcsSimpleSum << nl
-                << "    nonBlockingExchange: "
-                << UPstream::nProcsNonblockingExchange
-                << " (tuning: " << UPstream::tuning_NBX_ << ')' << nl
-                << "    exchange algorithm : "
-                << PstreamBuffers::algorithm << nl
-                << "    commsType          : "
-                << UPstream::commsTypeNames[UPstream::defaultCommsType] << nl
-                << "    polling iterations : "
-                << UPstream::nPollProcInterfaces << nl;
+            if (UPstream::floatTransfer)
+            {
+                Info<< "    floatTransfer      : enabled" << nl;
+            }
+            if (UPstream::maxCommsSize)
+            {
+                Info<< "    maxCommsSize       : "
+                    << UPstream::maxCommsSize << nl;
+            }
+            if (UPstream::nProcsSimpleSum > 2)
+            {
+                Info<< "    nProcsSimpleSum    : "
+                    << UPstream::nProcsSimpleSum << nl;
+            }
+            {
+                const auto& commsType =
+                    UPstream::commsTypeNames[UPstream::defaultCommsType];
+
+                Info<< "    nonBlockingExchange: "
+                    << UPstream::nProcsNonblockingExchange
+                    << " (tuning: " << UPstream::tuning_NBX_ << ')' << nl
+                    << "    exchange algorithm : "
+                    << PstreamBuffers::algorithm << nl
+                    << "    commsType          : " << commsType << nl
+                    << "    polling iterations : "
+                    << UPstream::nPollProcInterfaces << nl;
+            }
 
             if (UPstream::allWorlds().size() > 1)
             {