[Cherry-pick] NVLS support for NCCL API (#410) (#425)

Co-authored-by: Qinghua Zhou <[email protected]>
Co-authored-by: Changho Hwang <[email protected]>

Author: 3 people

apps/nccl/include/nccl.h

@@ -69,6 +69,15 @@ typedef struct ncclConfig_v21700 {
         NCCL_CONFIG_UNDEF_INT                             /* splitShare */     \
   }
 
+/* NCCL malloc and free function for all types of NCCL optimizations
+ * (e.g. user buffer registration). The actual allocated size might
+ * be larger than requested due to granularity requirement. */
+ncclResult_t ncclMemAlloc(void** ptr, size_t size);
+ncclResult_t pncclMemAlloc(void** ptr, size_t size);
+
+ncclResult_t ncclMemFree(void* ptr);
+ncclResult_t pncclMemFree(void* ptr);
+
 /* Return the NCCL_VERSION_CODE of the NCCL library in the supplied integer.
  * This integer is coded with the MAJOR, MINOR and PATCH level of the
  * NCCL library

apps/nccl/src/nccl.cu

@@ -8,6 +8,7 @@
 #include <mscclpp/executor.hpp>
 #include <mscclpp/sm_channel.hpp>
 #include <mscclpp/sm_channel_device.hpp>
+#include <mscclpp/utils.hpp>
 #include <sstream>
 #include <unordered_map>
 #include <vector>
@@ -33,6 +34,9 @@
 //                             mscclpp::Transport::IB3, mscclpp::Transport::IB4, mscclpp::Transport::IB5,
 //                             mscclpp::Transport::IB6, mscclpp::Transport::IB7};
 
+// Declare the global map to store associations between raw pointer and shared pointer
+static std::unordered_map<void*, std::shared_ptr<char>> ptrMap;
+
 struct channelKey {
   const void* buff;
   size_t bytes;
@@ -113,7 +117,7 @@ static size_t ncclTypeSize(ncclDataType_t type) {
   return 0;
 }
 
-double parseSize(const char* value) {
+static double parseSize(const char* value) {
   std::string valueStr(value);
   std::istringstream iss(valueStr);
   long long int units;
@@ -644,3 +648,59 @@ NCCL_API ncclResult_t ncclGroupEnd() {
   // Do nothing
   return ncclSuccess;
 }
+
+NCCL_API ncclResult_t ncclCommRegister(const ncclComm_t, void*, size_t, void**) {
+  // TODO: Implementation
+  return ncclSuccess;
+}
+
+NCCL_API ncclResult_t ncclCommDeregister(const ncclComm_t, void*) {
+  // TODO: Implementation
+  return ncclSuccess;
+}
+
+ncclResult_t ncclMemAlloc(void** ptr, size_t size) {
+  // Allocate memory using mscclpp::allocSharedPhysicalCuda
+  if (ptr == nullptr || size == 0) {
+    return ncclInvalidArgument;
+  }
+  std::shared_ptr<char> sharedPtr;
+  try {
+    if (mscclpp::isNvlsSupported()) {
+      sharedPtr = mscclpp::allocSharedPhysicalCuda<char>(size);
+    } else {
+      sharedPtr = mscclpp::allocExtSharedCuda<char>(size);
+    }
+    if (sharedPtr == nullptr) {
+      return ncclSystemError;
+    }
+  } catch (const mscclpp::Error& e) {
+    if (e.getErrorCode() == mscclpp::ErrorCode::InvalidUsage) {
+      return ncclInvalidUsage;
+    } else {
+      return ncclInternalError;
+    }
+  } catch (const mscclpp::CudaError& e) {
+    return ncclUnhandledCudaError;
+  } catch (const mscclpp::CuError& e) {
+    return ncclUnhandledCudaError;
+  } catch (const mscclpp::BaseError& e) {
+    return ncclInternalError;
+  }
+  ptrMap[sharedPtr.get()] = sharedPtr;
+
+  // Return the pointer
+  *ptr = sharedPtr.get();
+  return ncclSuccess;
+}
+
+ncclResult_t ncclMemFree(void* ptr) {
+  auto ptrIt = ptrMap.find(ptr);
+  if (ptrIt != ptrMap.end()) {
+    ptrMap.erase(ptrIt);
+    return ncclSuccess;
+  }
+
+  // Pointer not found
+  return ncclInvalidUsage;
+}

src/executor/execution_plan.cc

@@ -141,7 +141,17 @@ std::vector<ChannelInfo> ExecutionPlan::Impl::getUnpairedChannelInfos(int rank,
   return unpaired;
 }
 
-std::vector<NvlsInfo> ExecutionPlan::Impl::getNvlsInfos(int rank) const { return this->nvlsInfos.at(rank); }
+std::vector<NvlsInfo> ExecutionPlan::Impl::getNvlsInfos(int rank, size_t sendBuffserSize, size_t recvBufferSize) const {
+  if (sendBuffserSize == 0 && recvBufferSize == 0) {
+    return this->nvlsInfos.at(rank);
+  }
+  size_t chunkSize = this->getUpperBoundChunkSize(rank, sendBuffserSize, recvBufferSize);
+  std::vector<NvlsInfo> infos = this->nvlsInfos.at(rank);
+  for (auto& info : infos) {
+    info.bufferSize = info.bufferSize * chunkSize;
+  }
+  return infos;
+}
 
 std::vector<int> ExecutionPlan::Impl::getConnectedPeers(int rank) const {
   std::set<int> peers;
@@ -272,7 +282,7 @@ void ExecutionPlan::Impl::parseChannels(
       NvlsInfo info;
       info.bufferType = convertToBufferType(channel["buff"]);
       for (const auto& group : channel["rankGroups"]) {
-        info.bufferSize = (int)group["size"] * this->getUpperBoundChunkSize(rank, this->inputSize, this->outputSize);
+        info.bufferSize = (int)group["size"];
         info.ranks.clear();
         for (int rank : group["ranks"]) {
           info.ranks.push_back(rank);

src/executor/executor.cc

@@ -180,10 +180,10 @@ struct Executor::Impl {
     context.scratchBufferSize = scratchBufferSize;
     context.proxyService = std::make_shared<ProxyService>();
     context.nthreadsPerBlock = plan.impl_->getNThreadsPerBlock();
-    this->setupConnections(context, rank, plan);
+    this->setupConnections(context, rank, plan, sendMemRange, recvMemRange);
     this->setupRegisteredMemories(context, sendbuff, recvbuff, sendMemRange, recvMemRange, rank, plan);
     this->setupChannels(context, sendbuff, recvbuff, sendMemRange, recvMemRange, rank, plan);
-    this->setupNvlsChannels(context, sendbuff, recvbuff, rank, plan);
+    this->setupNvlsChannels(context, sendbuff, recvbuff, sendMemRange, recvMemRange, rank, plan);
     this->setupDeviceExecutionPlan(context, devicePlanKey, rank, plan);
     context.deviceExecutionPlansBuffers[devicePlanKey] =
         allocExtSharedCuda<char>(context.deviceExecutionPlans[devicePlanKey].size() * sizeof(DeviceExecutionPlan));
@@ -214,7 +214,8 @@ struct Executor::Impl {
     return flags;
   };
 
-  void setupConnections(ExecutionContext& context, int rank, const ExecutionPlan& plan) {
+  void setupConnections(ExecutionContext& context, int rank, const ExecutionPlan& plan, size_t sendBufferSize,
+                        size_t recvBufferSize) {
     std::vector<int> connectedPeers = plan.impl_->getConnectedPeers(rank);
     std::vector<mscclpp::NonblockingFuture<std::shared_ptr<mscclpp::Connection>>> connectionFutures;
     for (int peer : connectedPeers) {
@@ -227,7 +228,7 @@ struct Executor::Impl {
       context.connections[connectedPeers[i]] = connectionFutures[i].get();
     }
 
-    std::vector<NvlsInfo> nvlsInfos = plan.impl_->getNvlsInfos(rank);
+    std::vector<NvlsInfo> nvlsInfos = plan.impl_->getNvlsInfos(rank, sendBufferSize, recvBufferSize);
     for (const NvlsInfo& info : nvlsInfos) {
       std::shared_ptr<NvlsConnection> nvlsConnection =
           mscclpp::connectNvlsCollective(this->comm, info.ranks, info.bufferSize);
@@ -351,9 +352,9 @@ struct Executor::Impl {
     }
   }
 
-  void setupNvlsChannels(ExecutionContext& context, void* sendbuff, void* recvbuff, int rank,
-                         const ExecutionPlan& plan) {
-    std::vector<NvlsInfo> nvlsInfos = plan.impl_->getNvlsInfos(rank);
+  void setupNvlsChannels(ExecutionContext& context, void* sendbuff, void* recvbuff, size_t sendBufferSize,
+                         size_t recvBufferSize, int rank, const ExecutionPlan& plan) {
+    std::vector<NvlsInfo> nvlsInfos = plan.impl_->getNvlsInfos(rank, sendBufferSize, recvBufferSize);
     for (size_t i = 0; i < nvlsInfos.size(); i++) {
       std::shared_ptr<NvlsConnection> nvlsConnection = context.nvlsConnections[i];
       NvlsInfo info = nvlsInfos[i];

src/include/execution_plan.hpp

@@ -69,7 +69,7 @@ struct ExecutionPlan::Impl {
   std::vector<ChannelInfo> getChannelInfos(int rank, BufferType bufferType) const;
   std::vector<ChannelInfo> getChannelInfosByDstRank(int rank, BufferType bufferType) const;
   std::vector<ChannelInfo> getUnpairedChannelInfos(int rank, int worldSize, ChannelType channelType);
-  std::vector<NvlsInfo> getNvlsInfos(int rank) const;
+  std::vector<NvlsInfo> getNvlsInfos(int rank, size_t sendBuffserSize = 0, size_t recvBufferSize = 0) const;
   std::vector<int> getConnectedPeers(int rank) const;
   std::vector<BufferType> getConnectedBufferTypes(int rank) const;
   size_t getScratchBufferSize(int rank, size_t inputSize, size_t outputSize) const;