Map Exchange to set up communication for MDLCSR (#22)

* add mapping exchange test * remove map exchange test * mapping exchange implementation and test * increase memory and heap size * change main memory size from 16 GB back to 8 GB * fix cmake test script path * fix cmake test script path for pando-rt * remove comment
utcs-scea · Apr 9, 2024 · 9fbe68a · 9fbe68a
1 parent 2a33ec2
commit 9fbe68a
Show file tree

Hide file tree

Showing 8 changed files with 206 additions and 94 deletions.
diff --git a/.github/workflows/docker.yml b/.github/workflows/docker.yml
@@ -56,11 +56,11 @@ jobs:
         echo "IMAGE_VERSION=$(git log --pretty="%h" -1 Dockerfile.dev)" >> $GITHUB_ENV
         if [ ${{ matrix.build-type }} == 'LSAN' ]; then
           echo "PANDO_BUILD_DOCS=OFF" >> $GITHUB_ENV
-          echo "PANDO_CONTAINER_ENV=-e=PANDO_PREP_L1SP_HART=16384 -ePANDO_PREP_MAIN_NODE=8589934592" >> $GITHUB_ENV
+          echo "PANDO_CONTAINER_ENV=-e=PANDO_PREP_L1SP_HART=16384" >> $GITHUB_ENV
         fi
         if [ ${{ matrix.build-type }} == 'UBSAN' ]; then
           echo "PANDO_BUILD_DOCS=OFF" >> $GITHUB_ENV
-          echo "PANDO_CONTAINER_ENV=-e=PANDO_PREP_L1SP_HART=16384 -ePANDO_PREP_MAIN_NODE=8589934592" >> $GITHUB_ENV
+          echo "PANDO_CONTAINER_ENV=-e=PANDO_PREP_L1SP_HART=16384" >> $GITHUB_ENV
         fi
         if [ ${{ matrix.build-type }} == 'Release' ]; then
           echo "PANDO_BUILD_DOCS=OFF" >> $GITHUB_ENV

diff --git a/cmake/PANDOTesting.cmake b/cmake/PANDOTesting.cmake
@@ -69,9 +69,9 @@ function(pando_add_driver_test_lib TARGET SOURCEFILE LIBRARY)
   else ()
     set(HTHREADS "")
     if (${GASNet_CONDUIT} STREQUAL "smp")
-      set(DRIVER_SCRIPT ${pando-lib-galois_SOURCE_DIR}/pando-rt/scripts/preprun.sh)
+      set(DRIVER_SCRIPT ${pando-lib-galois_SOURCE_DIR}/scripts/preprun.sh)
     elseif (${GASNet_CONDUIT} STREQUAL "mpi")
-      set(DRIVER_SCRIPT ${pando-lib-galois_SOURCE_DIR}/pando-rt/scripts/preprun_mpi.sh)
+      set(DRIVER_SCRIPT ${pando-lib-galois_SOURCE_DIR}/scripts/preprun_mpi.sh)
     else ()
       message(FATAL_ERROR "No runner script for GASNet conduit ${GASNet_CONDUIT}")
     endif ()
@@ -124,9 +124,9 @@ endfunction()
 function(pando_add_bin_test TARGET ARGS INPUTFILE OKFILE)
   if (NOT PANDO_RT_BACKEND STREQUAL "DRVX")
     if (${GASNet_CONDUIT} STREQUAL "smp")
-      set(DRIVER_SCRIPT ${pando-lib-galois_SOURCE_DIR}/pando-rt/scripts/preprun.sh)
+      set(DRIVER_SCRIPT ${pando-lib-galois_SOURCE_DIR}/scripts/preprun.sh)
     elseif (${GASNet_CONDUIT} STREQUAL "mpi")
-      set(DRIVER_SCRIPT ${pando-lib-galois_SOURCE_DIR}/pando-rt/scripts/preprun_mpi.sh)
+      set(DRIVER_SCRIPT ${pando-lib-galois_SOURCE_DIR}/scripts/preprun_mpi.sh)
     else ()
       message(FATAL_ERROR "No runner script for GASNet conduit ${GASNet_CONDUIT}")
     endif ()

diff --git a/include/pando-lib-galois/graphs/dist_local_csr.hpp b/include/pando-lib-galois/graphs/dist_local_csr.hpp
@@ -487,10 +487,14 @@ class DistLocalCSR {
   }
 
   /** Host Information **/
-  std::uint64_t getPhysicalHostID(VertexTokenID tid) {
+  std::uint64_t getVirtualHostID(VertexTokenID tid) {
     std::uint64_t virtualHostID = tid % this->numVHosts();
-    std::uint64_t physicalHost = fmap(virtualToPhysicalMap.getLocalRef(), get, virtualHostID);
-    return physicalHost;
+    return virtualHostID;
+  }
+  std::uint64_t getPhysicalHostID(VertexTokenID tid) {
+    std::uint64_t virtualHostID = this->getVirtualHostID(tid);
+    std::uint64_t physicalHostID = fmap(virtualToPhysicalMap.getLocalRef(), get, virtualHostID);
+    return physicalHostID;
   }
 
   /** Topology Modifications **/

diff --git a/include/pando-lib-galois/graphs/mirror_dist_local_csr.hpp b/include/pando-lib-galois/graphs/mirror_dist_local_csr.hpp
@@ -8,7 +8,6 @@
 
 #include <utility>
 
-#include "pando-rt/sync/mutex.hpp"
 #include <pando-lib-galois/containers/hashtable.hpp>
 #include <pando-lib-galois/containers/host_indexed_map.hpp>
 #include <pando-lib-galois/containers/host_local_storage.hpp>
@@ -17,7 +16,10 @@
 #include <pando-lib-galois/graphs/local_csr.hpp>
 #include <pando-lib-galois/import/wmd_graph_importer.hpp>
 #include <pando-lib-galois/loops/do_all.hpp>
+#include <pando-lib-galois/sync/global_barrier.hpp>
+#include <pando-lib-galois/sync/simple_lock.hpp>
 #include <pando-lib-galois/utility/gptr_monad.hpp>
+#include <pando-lib-galois/utility/tuple.hpp>
 #include <pando-rt/containers/array.hpp>
 #include <pando-rt/containers/vector.hpp>
 #include <pando-rt/memory/memory_guard.hpp>
@@ -132,6 +134,10 @@ class MirrorDistLocalCSR {
     VertexTopologyID getMaster() {
       return master;
     }
+
+    bool operator==(const MirrorToMasterMap& a) noexcept {
+      return a.mirror == mirror && a.master == master;
+    }
   };
 
   /** Vertex Manipulation **/
@@ -239,19 +245,13 @@ class MirrorDistLocalCSR {
   }
 
   /** Host Information **/
+  std::uint64_t getVirtualHostID(VertexTokenID tid) {
+    return dlcsr.getVirtualHostID(tid);
+  }
   std::uint64_t getPhysicalHostID(VertexTokenID tid) {
     return dlcsr.getPhysicalHostID(tid);
   }
 
-  /** Sync **/
-  // TODO(Ying-Wei):
-  // write a sync function that reduces mirror values and then broadcasts master values
-  // return a bitmap of modified vertices
-  //
-  // template <typename Func>
-  // pando::Array<bool> sync(Func func, pando::Array<bool>) {
-  //}
-
   /**
    * @brief get vertex local dense ID
    */
@@ -303,8 +303,6 @@ class MirrorDistLocalCSR {
     return dlcsr.getLocalCSR();
   }
 
-  // TODO(Jeageun):
-  // write a initialize function that calls initializeAfterGather function of DistLocalCSR dlcsr
   template <typename ReadVertexType, typename ReadEdgeType>
   pando::Status initializeAfterGather(
       galois::HostLocalStorage<pando::Vector<ReadVertexType>> vertexData, std::uint64_t numVertices,
@@ -390,62 +388,81 @@ class MirrorDistLocalCSR {
       numVertices += lift(mirrorList[i], size);
     }
     PANDO_CHECK(wg.wait());
+
+    PANDO_CHECK_RETURN(setupCommunication());
+
     return pando::Status::Success;
   }
 
-  // TODO(Ying-Wei):
-  // uses doAll to send remoteMasterToLocalMirrorMap to corresponding remote hosts
-  // no need to use executeON
-  // just push to the localMasterToRemoteMirrorOrderedTable vector
-  // make sure to use the spin lock in pando-rt
   /**
-   * @brief Get the local mutex
+   * @brief Exchanges the mirror to master mapping from the mirror side to the maser side
    */
-  pando::GlobalRef<pando::Mutex> getLocalMutex(std::uint64_t host_id) {
-    return hostMutex[host_id];
-  }
-
   pando::Status setupCommunication() {
+    auto dims = pando::getPlaceDims();
+
+    // initialize localMirrorToRemoteMasterOrderedTable
     PANDO_CHECK_RETURN(localMasterToRemoteMirrorTable.initialize());
+    for (std::int16_t i = 0; i < dims.node.id; i++) {
+      pando::GlobalRef<pando::Vector<pando::Vector<MirrorToMasterMap>>>
+          localMasterToRemoteMirrorMap = localMasterToRemoteMirrorTable[i];
+      PANDO_CHECK_RETURN(fmap(localMasterToRemoteMirrorMap, initialize, dims.node.id));
+      for (std::int16_t i = 0; i < dims.node.id; i++) {
+        pando::GlobalRef<pando::Vector<MirrorToMasterMap>> mapVectorFromHost =
+            fmap(localMasterToRemoteMirrorMap, get, i);
+        PANDO_CHECK_RETURN(fmap(mapVectorFromHost, initialize, 0));
+      }
+    }
 
-    PANDO_CHECK_RETURN(hostMutex.initialize());
-
-    PANDO_CHECK_RETURN(galois::doAll(
-        localMirrorToRemoteMasterOrderedTable, localMasterToRemoteMirrorTable,
-        +[](galois::HostLocalStorage<pando::Array<MirrorToMasterMap>>
-                localMirrorToRemoteMasterOrderedTable,
-            pando::GlobalRef<pando::Vector<EdgeHandle>> localMasterToRemoteMirrorTable) {
-          PANDO_CHECK(fmap(localMirrorToRemoteMasterOrderedTable, initialize, 0));
-          pando::Array<MirrorToMasterMap> remoteMasterToLocalMirrorMap =
-              localMirrorToRemoteMasterOrderedTable.getLocal();
-          for (MirrorToMasterMap m : remoteMasterToLocalMirrorMap) {
-            VertexTopologyID masterTopologyID = m.master;
-            VertexTokenID masterTokenID = getTokenID(masterTopologyID);
-            std::uint64_t physicalHost = getPhysicalHostID(masterTokenID);
-            pando::Mutex mutex = getLocalMutex(physicalHost);
-
-            // Lock mutex to ensure atomic append to the vector
-            mutex.lock();
-            PANDO_CHECK(fmap(localMasterToRemoteMirrorTable, pushBack, m));
-            mutex.unlock();
+    auto thisCSR = *this;
+    auto state = galois::make_tpl(thisCSR, localMasterToRemoteMirrorTable);
+
+    // push style
+    // each host traverses its own localMirrorToRemoteMasterOrderedTable and send out the mapping to
+    // the corresponding remote host append to the vector of vector where each vector is the mapping
+    // from a specific host
+    galois::doAll(
+        state, localMirrorToRemoteMasterOrderedTable,
+        +[](decltype(state) state,
+            pando::GlobalRef<pando::Array<MirrorToMasterMap>> localMirrorToRemoteMasterOrderedMap) {
+          auto [object, localMasterToRemoteMirrorTable] = state;
+          for (std::uint64_t i = 0ul; i < lift(localMirrorToRemoteMasterOrderedMap, size); i++) {
+            MirrorToMasterMap m = fmap(localMirrorToRemoteMasterOrderedMap, get, i);
+            VertexTopologyID masterTopologyID = m.getMaster();
+            VertexTokenID masterTokenID = object.getTokenID(masterTopologyID);
+            std::uint64_t physicalHost = object.getPhysicalHostID(masterTokenID);
+
+            pando::GlobalRef<pando::Vector<pando::Vector<MirrorToMasterMap>>>
+                localMasterToRemoteMirrorMap = localMasterToRemoteMirrorTable[physicalHost];
+            pando::GlobalRef<pando::Vector<MirrorToMasterMap>> mapVectorFromHost =
+                fmap(localMasterToRemoteMirrorMap, get, pando::getCurrentPlace().node.id);
+
+            PANDO_CHECK(fmap(mapVectorFromHost, pushBack, m));
           }
-        }));
+        });
 
     return pando::Status::Success;
   }
 
+  /**
+   * @brief For testing only
+   */
+  pando::GlobalRef<pando::Array<MirrorToMasterMap>> getLocalMirrorToRemoteMasterOrderedMap(
+      int16_t hostId) {
+    return localMirrorToRemoteMasterOrderedTable[hostId];
+  }
+  pando::GlobalRef<pando::Vector<pando::Vector<MirrorToMasterMap>>> getLocalMasterToRemoteMirrorMap(
+      uint64_t hostId) {
+    return localMasterToRemoteMirrorTable[hostId];
+  }
+
 private:
   DLCSR dlcsr;
   uint64_t _mirror_size;
   galois::HostLocalStorage<LocalVertexRange> masterRange;
   galois::HostLocalStorage<LocalVertexRange> mirrorRange;
   galois::HostLocalStorage<pando::Array<MirrorToMasterMap>> localMirrorToRemoteMasterOrderedTable;
-
-  // TODO(Ying-Wei):
-  // generate the following
-  galois::HostLocalStorage<pando::Mutex> hostMutex;
-  galois::HostLocalStorage<pando::Vector<EdgeHandle>> localMasterToRemoteMirrorTable;
-  // galois::GlobalBarrier barrier;
+  galois::HostLocalStorage<pando::Vector<pando::Vector<MirrorToMasterMap>>>
+      localMasterToRemoteMirrorTable;
 };
 
 static_assert(graph_checker<MirrorDistLocalCSR<std::uint64_t, std::uint64_t>>::value);

diff --git a/pando-rt/cmake/PANDOTesting.cmake b/pando-rt/cmake/PANDOTesting.cmake
@@ -72,9 +72,9 @@ function(pando_add_driver_test TARGET SOURCEFILE)
 
   if (PANDO_RT_BACKEND STREQUAL "PREP")
     if (${GASNet_CONDUIT} STREQUAL "smp")
-      set(RUNNER_SCRIPT ${PROJECT_SOURCE_DIR}/scripts/preprun.sh)
+      set(RUNNER_SCRIPT ${PROJECT_SOURCE_DIR}/../scripts/preprun.sh)
     elseif (${GASNet_CONDUIT} STREQUAL "mpi")
-      set(RUNNER_SCRIPT ${PROJECT_SOURCE_DIR}/scripts/preprun_mpi.sh)
+      set(RUNNER_SCRIPT ${PROJECT_SOURCE_DIR}/../scripts/preprun_mpi.sh)
     else ()
       message(FATAL_ERROR "No runner script for GASNet conduit ${GASNet_CONDUIT}")
     endif ()

diff --git a/pando-rt/scripts/preprun.sh → scripts/preprun.sh b/pando-rt/scripts/preprun.sh → scripts/preprun.sh
@@ -1,4 +1,7 @@
 #!/bin/bash
+# SPDX-License-Identifier: MIT
+# Copyright (c) 2023. University of Texas at Austin. All rights reserved.
+
 #
 # SPDX-License-Identifier: MIT
 # Copyright (c) 2023 Advanced Micro Devices, Inc. All rights reserved.
@@ -29,32 +32,31 @@ usage: preprun -n <n> prog [program args]
 
 harts=""
 
-while getopts "n:c:t:h" option
-do
+while getopts "n:c:t:h" option; do
     case ${option} in
-        n) # number of emulated PXNs
-            nodes=${OPTARG}
-            ;;
-        c) # number of emulated cores per PXN
-            cores=${OPTARG}
-            ;;
-        t) # number of emulated cores per PXN
-            harts=${OPTARG}
-            ;;
-        h) # help
-            show_help
-            exit
-            ;;
+    n) # number of emulated PXNs
+        nodes=${OPTARG}
+        ;;
+    c) # number of emulated cores per PXN
+        cores=${OPTARG}
+        ;;
+    t) # number of emulated cores per PXN
+        harts=${OPTARG}
+        ;;
+    h) # help
+        show_help
+        exit
+        ;;
     esac
 done
-shift $(expr $OPTIND - 1 )
+shift $(expr $OPTIND - 1)
 prog=$@
 
 export GASNET_PSHM_NODES=$nodes
 export PANDO_PREP_NUM_CORES=$cores
 
 if [ -n "$harts" ]; then
-  export PANDO_PREP_NUM_HARTS=$harts
+    export PANDO_PREP_NUM_HARTS=$harts
 fi
 
 exec $prog
diff --git a/pando-rt/scripts/preprun_mpi.sh → scripts/preprun_mpi.sh b/pando-rt/scripts/preprun_mpi.sh → scripts/preprun_mpi.sh
@@ -1,4 +1,7 @@
 #!/bin/bash
+# SPDX-License-Identifier: MIT
+# Copyright (c) 2023. University of Texas at Austin. All rights reserved.
+
 #
 # SPDX-License-Identifier: MIT
 # Copyright (c) 2023 Advanced Micro Devices, Inc. All rights reserved.
@@ -28,32 +31,35 @@ usage: preprun -n <n> prog [program args]
 #
 
 harts=""
+# 16GB Main memory size by default
+main_memory_size="${main_memory_size-8589934592}"
 
-while getopts "n:c:t:h" option
-do
+while getopts "n:c:t:h" option; do
     case ${option} in
-        n) # number of emulated PXNs
-            nodes=${OPTARG}
-            ;;
-        c) # number of emulated cores per PXN
-            cores=${OPTARG}
-            ;;
-        t) # number of emulated cores per PXN
-            harts=${OPTARG}
-            ;;
-        h) # help
-            show_help
-            exit
-            ;;
+    n) # number of emulated PXNs
+        nodes=${OPTARG}
+        ;;
+    c) # number of emulated cores per PXN
+        cores=${OPTARG}
+        ;;
+    t) # number of emulated cores per PXN
+        harts=${OPTARG}
+        ;;
+    h) # help
+        show_help
+        exit
+        ;;
     esac
 done
-shift $(expr $OPTIND - 1 )
+shift $(expr $OPTIND - 1)
 prog=$@
 
 export PANDO_PREP_NUM_CORES=$cores
 
 if [ -n "$harts" ]; then
-  export PANDO_PREP_NUM_HARTS=$harts
+    export PANDO_PREP_NUM_HARTS=$harts
 fi
 
+export PANDO_PREP_MAIN_NODE=$main_memory_size
+
 gasnetrun_mpi -n $nodes $prog