Merge pull request #66 from microsoft/olli/api-extension

Olli/api extension

CMakeLists.txt

@@ -0,0 +1,32 @@
+cmake_minimum_required(VERSION 3.26)
+
+project(mscclpp LANGUAGES CUDA CXX)
+
+set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CUDA_STANDARD 17)
+
+list(APPEND CMAKE_MODULE_PATH ${PROJECT_SOURCE_DIR}/cmake/modules)
+
+find_package(CUDAToolkit REQUIRED)
+find_package(IBVerbs REQUIRED)
+find_package(NUMA REQUIRED)
+find_package(GDRCopy)
+
+option(USE_MPI_FOR_TESTS "Use MPI for tests" ON)
+if(USE_MPI_FOR_TESTS)
+    find_package(MPI REQUIRED)
+    add_definitions(-DMSCCLPP_USE_MPI_FOR_TESTS)
+endif()
+
+include_directories(${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
+
+add_library(mscclpp SHARED)
+add_subdirectory(src) # This adds the srouces to the mscclpp target
+target_include_directories(mscclpp PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/src/include)
+set_target_properties(mscclpp PROPERTIES LINKER_LANGUAGE CXX)
+target_link_libraries(mscclpp PRIVATE MSCCLPP::ibverbs MSCCLPP::numa CUDA::cudart CUDA::cuda_driver)
+if(GDRCOPY_FOUND)
+    target_link_libraries(mscclpp PRIVATE MSCCLPP::gdrcopy)
+endif()
+
+add_subdirectory(tests)

Makefile

@@ -61,7 +61,7 @@ endif
 
 NVCUFLAGS  := -ccbin $(CXX) $(NVCC_GENCODE) -std=c++11 --expt-extended-lambda -Xfatbin -compress-all
 # Use addprefix so that we can specify more than one path
-NVLDFLAGS  := -L$(CUDA_LIB) -lcudart -lrt
+NVLDFLAGS  := -L$(CUDA_LIB) -lcudart -lrt -lcuda
 
 ifeq ($(DEBUG), 0)
 NVCUFLAGS += -O3
@@ -120,7 +120,8 @@ LDFLAGS := $(NVLDFLAGS) $(GDRCOPY_LDFLAGS) -libverbs -lnuma
 
 LIBSRCS := $(addprefix src/,debug.cc utils.cc init.cc proxy.cc ib.cc config.cc)
 LIBSRCS += $(addprefix src/bootstrap/,bootstrap.cc socket.cc)
-LIBSRCS += $(addprefix src/,communicator.cc fifo.cc host_connection.cc proxy_cpp.cc basic_proxy_handler.cc)
+LIBSRCS += $(addprefix src/,communicator.cc connection.cc registered_memory.cc)
+LIBSRCS += $(addprefix src/,epoch.cc proxy_cpp.cc fifo.cc channel.cc)
 ifneq ($(NPKIT), 0)
 LIBSRCS += $(addprefix src/misc/,npkit.cc)
 endif
@@ -134,7 +135,7 @@ HEADERS := $(wildcard src/include/*.h)
 CPPSOURCES := $(shell find ./ -regextype posix-extended -regex '.*\.(c|cpp|h|hpp|cc|cxx|cu)' -not -path "./build/*" -not -path "./python/*")
 PYTHONCPPSOURCES := $(shell find ./python/src/ -regextype posix-extended -regex '.*\.(c|cpp|h|hpp|cc|cxx|cu)')
 
-INCEXPORTS := mscclpp.h mscclppfifo.h mscclpp.hpp mscclppfifo.hpp
+INCEXPORTS := mscclpp.h mscclppfifo.h mscclpp.hpp mscclppfifo.hpp epoch.hpp
 INCTARGETS := $(INCEXPORTS:%=$(BUILDDIR)/$(INCDIR)/%)
 
 LIBNAME   := libmscclpp.so
@@ -148,7 +149,7 @@ UTOBJTARGETS := $(UTOBJS:%=$(BUILDDIR)/$(OBJDIR)/%)
 UTBINS       := $(patsubst %.o,$(BUILDDIR)/$(BINDIR)/%,$(UTOBJS))
 
 TESTSDIR  := tests
-TESTSSRCS := $(addprefix $(TESTSDIR)/,bootstrap_test.cc allgather_test_standalone.cu allgather_test_cpp.cu bootstrap_test_cpp.cc)
+TESTSSRCS := $(addprefix $(TESTSDIR)/,bootstrap_test.cc allgather_test_standalone.cu  communicator_test_cpp.cu bootstrap_test_cpp.cc allgather_test_cpp.cu)
 TESTSOBJS := $(patsubst %.cc,%.o,$(TESTSSRCS)) $(patsubst %.cu,%.o,$(TESTSSRCS))
 TESTSOBJTARGETS := $(TESTSOBJS:%=$(BUILDDIR)/$(OBJDIR)/%)
 TESTSBINS       := $(patsubst %.o,$(BUILDDIR)/$(BINDIR)/%,$(TESTSOBJS))

TODO.md

@@ -0,0 +1,8 @@
+# Core API extraction
+
+- Add a test for host side Communicator/RegisteredMemory/Connection use.
+- Implement a standalone "epoch" synchronization construct that can be used as a component in custom proxies. epoch.hpp/cc has the beginnings of this.
+- Reimplement the "standard" proxy service + DeviceConnection on top of the new Communicator/RegisteredMemory/Connection core API. Remants of the old code is in channel.hpp, basic_proxy_handler.hpp/cc and host_connection.hpp/cc. Probably need a manager class to wrap all of this.
+- Change the new IBConnection and Communicator to use the new C++ IbCtx and IbQp classes.
+- Implement IbQp::~IbQp()
+- Fix RegisteredMemory::Impl::Impl to get the IPC handle from the base pointer, not the derived pointer.

cmake/modules/FindGDRCopy.cmake

@@ -0,0 +1,41 @@
+# Find the GDRCopy libraries
+#
+# The following variables are optionally searched for defaults
+#  GDRCOPY_ROOT_DIR: Base directory where all GDRCopy components are found
+#  GDRCOPY_INCLUDE_DIR: Directory where GDRCopy headers are found
+#  GDRCOPY_LIB_DIR: Directory where GDRCopy libraries are found
+
+# The following are set after configuration is done:
+#  GDRCOPY_FOUND
+#  GDRCOPY_INCLUDE_DIRS
+#  GDRCOPY_LIBRARIES
+
+# An imported target MSCCLPP::gdrcopy is created if the library is found.
+
+find_path(GDRCOPY_INCLUDE_DIRS
+  NAMES gdrapi.h
+  HINTS
+  ${GDRCOPY_INCLUDE_DIR}
+  ${GDRCOPY_ROOT_DIR}
+  ${GDRCOPY_ROOT_DIR}/include)
+
+find_library(GDRCOPY_LIBRARIES
+  NAMES gdrapi
+  HINTS
+  ${GDRCOPY_LIB_DIR}
+  ${GDRCOPY_ROOT_DIR}
+  ${GDRCOPY_ROOT_DIR}/lib)
+
+include(FindPackageHandleStandardArgs)
+find_package_handle_standard_args(GDRCopy DEFAULT_MSG GDRCOPY_INCLUDE_DIRS GDRCOPY_LIBRARIES)
+mark_as_advanced(GDRCOPY_INCLUDE_DIR GDRCOPY_LIBRARIES)
+
+if(GDRCOPY_FOUND)
+  if(NOT TARGET MSCCLPP::gdrcopy)
+    add_library(MSCCLPP::gdrcopy UNKNOWN IMPORTED)
+  endif()
+  set_target_properties(MSCCLPP::gdrcopy PROPERTIES
+    INTERFACE_INCLUDE_DIRECTORIES "${GDRCOPY_INCLUDE_DIR}"
+    IMPORTED_LINK_INTERFACE_LANGUAGES "C"
+    IMPORTED_LOCATION "${GDRCOPY_LIBRARIES}")
+endif()

cmake/modules/FindIBVerbs.cmake

@@ -0,0 +1,41 @@
+# Find the IB Verbs libraries
+#
+# The following variables are optionally searched for defaults
+#  IBVERBS_ROOT_DIR: Base directory where all ibverbs components are found
+#  IBVERBS_INCLUDE_DIR: Directory where ibverbs headers are found
+#  IBVERBS_LIB_DIR: Directory where ibverbs libraries are found
+
+# The following are set after configuration is done:
+#  IBVERBS_FOUND
+#  IBVERBS_INCLUDE_DIRS
+#  IBVERBS_LIBRARIES
+
+# An imported target MSCCLPP::ibverbs is created if the library is found.
+
+find_path(IBVERBS_INCLUDE_DIRS
+  NAMES infiniband/verbs.h
+  HINTS
+  ${IBVERBS_INCLUDE_DIR}
+  ${IBVERBS_ROOT_DIR}
+  ${IBVERBS_ROOT_DIR}/include)
+
+find_library(IBVERBS_LIBRARIES
+  NAMES ibverbs
+  HINTS
+  ${IBVERBS_LIB_DIR}
+  ${IBVERBS_ROOT_DIR}
+  ${IBVERBS_ROOT_DIR}/lib)
+
+include(FindPackageHandleStandardArgs)
+find_package_handle_standard_args(IBVerbs DEFAULT_MSG IBVERBS_INCLUDE_DIRS IBVERBS_LIBRARIES)
+mark_as_advanced(IBVERBS_INCLUDE_DIR IBVERBS_LIBRARIES)
+
+if(IBVERBS_FOUND)
+  if(NOT TARGET MSCCLPP::ibverbs)
+    add_library(MSCCLPP::ibverbs UNKNOWN IMPORTED)
+  endif()
+  set_target_properties(MSCCLPP::ibverbs PROPERTIES
+    INTERFACE_INCLUDE_DIRECTORIES "${IBVERBS_INCLUDE_DIR}"
+    IMPORTED_LINK_INTERFACE_LANGUAGES "C"
+    IMPORTED_LOCATION "${IBVERBS_LIBRARIES}")
+endif()

cmake/modules/FindNUMA.cmake

@@ -0,0 +1,41 @@
+# Find the numa libraries
+#
+# The following variables are optionally searched for defaults
+#  NUMA_ROOT_DIR: Base directory where all numa components are found
+#  NUMA_INCLUDE_DIR: Directory where numa headers are found
+#  NUMA_LIB_DIR: Directory where numa libraries are found
+
+# The following are set after configuration is done:
+#  NUMA_FOUND
+#  NUMA_INCLUDE_DIRS
+#  NUMA_LIBRARIES
+
+# An imported target MSCCLPP::numa is created if the library is found.
+
+find_path(NUMA_INCLUDE_DIRS
+  NAMES numa.h
+  HINTS
+  ${NUMA_INCLUDE_DIR}
+  ${NUMA_ROOT_DIR}
+  ${NUMA_ROOT_DIR}/include)
+
+find_library(NUMA_LIBRARIES
+  NAMES numa
+  HINTS
+  ${NUMA_LIB_DIR}
+  ${NUMA_ROOT_DIR}
+  ${NUMA_ROOT_DIR}/lib)
+
+include(FindPackageHandleStandardArgs)
+find_package_handle_standard_args(NUMA DEFAULT_MSG NUMA_INCLUDE_DIRS NUMA_LIBRARIES)
+mark_as_advanced(NUMA_INCLUDE_DIR NUMA_LIBRARIES)
+
+if(NUMA_FOUND)
+  if(NOT TARGET MSCCLPP::numa)
+    add_library(MSCCLPP::numa UNKNOWN IMPORTED)
+  endif()
+  set_target_properties(MSCCLPP::numa PROPERTIES
+    INTERFACE_INCLUDE_DIRECTORIES "${NUMA_INCLUDE_DIR}"
+    IMPORTED_LINK_INTERFACE_LANGUAGES "C"
+    IMPORTED_LOCATION "${NUMA_LIBRARIES}")
+endif()

src/CMakeLists.txt

@@ -0,0 +1,5 @@
+file(GLOB_RECURSE SOURCES CONFIGURE_DEPENDS *.cc *.h)
+file(GLOB to_remove gdr.cc)
+list(REMOVE_ITEM SOURCES ${to_remove})
+
+target_sources(mscclpp PRIVATE ${SOURCES})

src/bootstrap/bootstrap.cc

@@ -180,9 +180,8 @@ Bootstrap::Impl::~Impl()
   }
 }
 
-void Bootstrap::Impl::getRemoteAddresses(mscclppSocket* listenSock,
-                                                std::vector<mscclppSocketAddress>& rankAddresses,
-                                                std::vector<mscclppSocketAddress>& rankAddressesRoot, int& rank)
+void Bootstrap::Impl::getRemoteAddresses(mscclppSocket* listenSock, std::vector<mscclppSocketAddress>& rankAddresses,
+                                         std::vector<mscclppSocketAddress>& rankAddressesRoot, int& rank)
 {
   mscclppSocket sock;
   ExtInfo info;
@@ -211,7 +210,7 @@ void Bootstrap::Impl::getRemoteAddresses(mscclppSocket* listenSock,
 }
 
 void Bootstrap::Impl::sendHandleToPeer(int peer, const std::vector<mscclppSocketAddress>& rankAddresses,
-                                              const std::vector<mscclppSocketAddress>& rankAddressesRoot)
+                                       const std::vector<mscclppSocketAddress>& rankAddressesRoot)
 {
   mscclppSocket sock;
   int next = (peer + 1) % this->nRanks_;
@@ -226,7 +225,8 @@ void Bootstrap::Impl::bootstrapCreateRoot()
   mscclppSocket listenSock;
 
   // mscclppSocket* listenSock = new mscclppSocket(); // TODO(saemal) make this a shared ptr
-  MSCCLPPTHROW(mscclppSocketInit(&listenSock, &uniqueId_.addr, uniqueId_.magic, mscclppSocketTypeBootstrap, nullptr, 0));
+  MSCCLPPTHROW(
+    mscclppSocketInit(&listenSock, &uniqueId_.addr, uniqueId_.magic, mscclppSocketTypeBootstrap, nullptr, 0));
   MSCCLPPTHROW(mscclppSocketListen(&listenSock));
   MSCCLPPTHROW(mscclppSocketGetAddr(&listenSock, &uniqueId_.addr));
   auto lambda = [this, listenSock]() { this->bootstrapRoot(listenSock); };

src/channel.cc

@@ -0,0 +1,26 @@
+#include "channel.hpp"
+#include "utils.h"
+#include "checks.hpp"
+#include "api.h"
+#include "debug.h"
+
+namespace mscclpp {
+namespace channel {
+
+MSCCLPP_API_CPP DeviceChannelService::DeviceChannelService(Communicator& communicator) : communicator_(communicator),
+    proxy_([&](ProxyTrigger triggerRaw) { return handleTrigger(triggerRaw); }, [&]() { bindThread(); }) {
+  int cudaDevice;
+  CUDATHROW(cudaGetDevice(&cudaDevice));
+  MSCCLPPTHROW(getDeviceNumaNode(cudaDevice, &deviceNumaNode));
+}
+
+MSCCLPP_API_CPP void DeviceChannelService::bindThread()
+{
+  if (deviceNumaNode >= 0) {
+    MSCCLPPTHROW(numaBind(deviceNumaNode));
+    INFO(MSCCLPP_INIT, "NUMA node of DeviceChannelService proxy thread is set to %d", deviceNumaNode);
+  }
+}
+
+} // namespace channel
+} // namespace mscclpp