mpi(hwloc): spin-lock receiver threads and pin to a core

Author: csegarragonz

.github/workflows/tests.yml

@@ -21,6 +21,8 @@ jobs:
     runs-on: ubuntu-latest
     container:
       image: faasm.azurecr.io/faabric:0.16.0
+    env:
+      FAABRIC_DEPLOYMENT_TYPE: gha-ci
     steps:
       - name: "Check-out code"
         uses: actions/checkout@v4
@@ -59,7 +61,7 @@ jobs:
     needs: [conan-cache]
     runs-on: ubuntu-latest
     env:
-      HOST_TYPE: ci
+      FAABRIC_DEPLOYMENT_TYPE: gha-ci
       REDIS_QUEUE_HOST: redis
       REDIS_STATE_HOST: redis
     container:
@@ -98,7 +100,7 @@ jobs:
       matrix:
         sanitiser: [None, Address, Thread, Undefined]
     env:
-      HOST_TYPE: ci
+      FAABRIC_DEPLOYMENT_TYPE: gha-ci
       REDIS_QUEUE_HOST: redis
       REDIS_STATE_HOST: redis
     container:
@@ -145,7 +147,7 @@ jobs:
     needs: [conan-cache]
     runs-on: ubuntu-latest
     env:
-      HOST_TYPE: ci
+      FAABRIC_DEPLOYMENT_TYPE: gha-ci
       REDIS_QUEUE_HOST: redis
       REDIS_STATE_HOST: redis
     container:

CMakeLists.txt

@@ -6,6 +6,7 @@ option(FAABRIC_BUILD_TESTS "Build Faabric tests" ON)
 option(FAABRIC_SELF_TRACING "Turn on system tracing using the logger" OFF)
 option(FAABRIC_CODE_COVERAGE "Build Faabric with code coverage profiling" OFF)
 option(FAABRIC_TARGET_CPU "CPU to optimise for, e.g. skylake, icelake or native" OFF)
+option(FAABRIC_USE_SPINLOCK "Use spinlocks for low-latency messaging" ON)
 
 # Enable colorized compiler output
 if("${CMAKE_CXX_COMPILER_ID}" STREQUAL "GNU")
@@ -38,6 +39,15 @@ if(${FAABRIC_SELF_TRACING})
     add_definitions(-DTRACE_ALL=1)
 endif()
 
+# We want to disable the usage of spinlocks (and CPU pinning) in GHA runners
+# as they have a very low number of vCPU cores
+if (${FAABRIC_USE_SPINLOCK} AND NOT ENV{DEPLOYMENT_TYPE} STREQUAL "gha-ci")
+    message(STATUS "Faabric: Activated spin-locks")
+    add_definitions(-DFAABRIC_USE_SPINLOCK)
+else()
+    message(STATUS "Faabric: Disabled spin-locks")
+endif()
+
 # Set-up use of sanitisers
 if (FAABRIC_USE_SANITISER STREQUAL "Address")
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsanitize=address")

cmake/ExternalProjects.cmake

@@ -15,6 +15,7 @@ if(NOT EXISTS "${CMAKE_CURRENT_BINARY_DIR}/conan.cmake")
                 TLS_VERIFY ON)
 endif()
 
+set(CONAN_CMAKE_SILENT_OUTPUT ON CACHE INTERNAL "")
 include(${CMAKE_CURRENT_BINARY_DIR}/conan.cmake)
 
 conan_check(VERSION 1.63.0 REQUIRED)
@@ -73,19 +74,19 @@ conan_cmake_install(PATH_OR_REFERENCE .
 
 include(${CMAKE_CURRENT_BINARY_DIR}/conan_paths.cmake)
 
-find_package(absl REQUIRED)
-find_package(Boost 1.80.0 REQUIRED)
-find_package(Catch2 REQUIRED)
-find_package(flatbuffers REQUIRED)
-find_package(fmt REQUIRED)
-find_package(hiredis REQUIRED)
+find_package(absl QUIET REQUIRED)
+find_package(Boost 1.80.0 QUIET REQUIRED)
+find_package(Catch2 QUIET REQUIRED)
+find_package(flatbuffers QUIET REQUIRED)
+find_package(fmt QUIET REQUIRED)
+find_package(hiredis QUIET REQUIRED)
 # 27/01/2023 - Pin OpenSSL to a specific version to avoid incompatibilities
 # with the system's (i.e. Ubuntu 22.04) OpenSSL
-find_package(OpenSSL 3.0.2 REQUIRED)
-find_package(Protobuf 3.20.0 REQUIRED)
-find_package(readerwriterqueue REQUIRED)
-find_package(spdlog REQUIRED)
-find_package(ZLIB REQUIRED)
+find_package(OpenSSL 3.0.2 QUIET REQUIRED)
+find_package(Protobuf 3.20.0 QUIET REQUIRED)
+find_package(readerwriterqueue QUIET REQUIRED)
+find_package(spdlog QUIET REQUIRED)
+find_package(ZLIB QUIET REQUIRED)
 
 # --------------------------------
 # Fetch content dependencies
@@ -109,16 +110,26 @@ set(ZSTD_LZ4_SUPPORT OFF CACHE INTERNAL "")
 # nng (Conan version out of date)
 set(NNG_TESTS OFF CACHE INTERNAL "")
 
-FetchContent_Declare(zstd_ext
-    GIT_REPOSITORY "https://github.com/facebook/zstd"
-    GIT_TAG "v1.5.2"
-    SOURCE_SUBDIR "build/cmake"
+FetchContent_Declare(atomic_queue_ext
+    GIT_REPOSITORY "https://github.com/max0x7ba/atomic_queue"
+    GIT_TAG "7c36f0997979a0fee5be84c9511ee0f6032057ec"
 )
 FetchContent_Declare(nng_ext
     GIT_REPOSITORY "https://github.com/nanomsg/nng"
     # NNG tagged version 1.7.1
     GIT_TAG "ec4b5722fba105e3b944e3dc0f6b63c941748b3f"
 )
+FetchContent_Declare(zstd_ext
+    GIT_REPOSITORY "https://github.com/facebook/zstd"
+    GIT_TAG "v1.5.2"
+    SOURCE_SUBDIR "build/cmake"
+)
+
+FetchContent_MakeAvailable(atomic_queue_ext)
+add_library(atomic_queue::atomic_queue ALIAS atomic_queue)
+
+FetchContent_MakeAvailable(nng_ext)
+add_library(nng::nng ALIAS nng)
 
 FetchContent_MakeAvailable(zstd_ext)
 # Work around zstd not declaring its targets properly
@@ -127,9 +138,6 @@ target_include_directories(libzstd_shared SYSTEM INTERFACE $<BUILD_INTERFACE:${z
 add_library(zstd::libzstd_static ALIAS libzstd_static)
 add_library(zstd::libzstd_shared ALIAS libzstd_shared)
 
-FetchContent_MakeAvailable(nng_ext)
-add_library(nng::nng ALIAS nng)
-
 # Group all external dependencies into a convenient virtual CMake library
 add_library(faabric_common_dependencies INTERFACE)
 target_include_directories(faabric_common_dependencies INTERFACE
@@ -141,6 +149,7 @@ target_link_libraries(faabric_common_dependencies INTERFACE
     absl::flat_hash_set
     absl::flat_hash_map
     absl::strings
+    atomic_queue::atomic_queue
     Boost::Boost
     Boost::system
     flatbuffers::flatbuffers

dist-test/build.sh

@@ -8,6 +8,7 @@ pushd ${PROJ_ROOT} >> /dev/null
 # Run the build
 docker compose \
     run \
+    -e FAABRIC_DEPLOYMENT_TYPE=gha-ci \
     --rm \
     cli \
     /code/faabric/dist-test/build_internal.sh

docker-compose.yml

@@ -29,6 +29,7 @@ services:
     tty: true
     privileged: true
     environment:
+      - DEPLOYMENT_TYPE=${FAABRIC_DEPLOYMENT_TYPE:-compose}
       - LOG_LEVEL=${LOG_LEVEL:-debug}
       - PLANNER_HOST=planner
       - PLANNER_PORT=8080
@@ -48,6 +49,7 @@ services:
       - ./conan-cache/:/root/.conan
     working_dir: /build/faabric/static
     environment:
+      - DEPLOYMENT_TYPE=${FAABRIC_DEPLOYMENT_TYPE:-compose}
       - LOG_LEVEL=debug
       - PLANNER_HOST=planner
       - PLANNER_PORT=8080

include/faabric/mpi/MpiWorld.h

@@ -26,7 +26,11 @@ namespace faabric::mpi {
 // as the broker already has mocking capabilities
 std::vector<MpiMessage> getMpiMockedMessages(int sendRank);
 
+#ifdef FAABRIC_USE_SPINLOCK
+typedef faabric::util::SpinLockQueue<MpiMessage> InMemoryMpiQueue;
+#else
 typedef faabric::util::FixedCapacityQueue<MpiMessage> InMemoryMpiQueue;
+#endif
 
 class MpiWorld
 {

include/faabric/transport/tcp/RecvSocket.h

@@ -31,5 +31,7 @@ class RecvSocket
     int port;
 
     std::deque<int> openConnections;
+
+    void setSocketOptions(int connFd);
 };
 }

include/faabric/transport/tcp/SocketOptions.h

@@ -5,8 +5,11 @@ void setReuseAddr(int connFd);
 void setNoDelay(int connFd);
 void setQuickAck(int connFd);
 
+// Blocking/Non-blocking sockets
 void setNonBlocking(int connFd);
 void setBlocking(int connFd);
-
 bool isNonBlocking(int connFd);
+
+// Enable busy polling for non-blocking sockets
+void setBusyPolling(int connFd);
 }

include/faabric/util/config.h

@@ -27,6 +27,7 @@ class SystemConfig
 
     // Scheduling
     int overrideCpuCount;
+    int overrideFreeCpuStart;
     std::string batchSchedulerMode;
 
     // Worker-related timeouts

include/faabric/util/hwloc.h

@@ -0,0 +1,32 @@
+#pragma once
+
+#include <memory>
+#include <pthread.h>
+
+namespace faabric::util {
+
+const int NO_CPU_IDX = -1;
+const int GHA_CPU_IDX = -2;
+
+class FaabricCpuSet
+{
+  public:
+    FaabricCpuSet(int cpuIdxIn = NO_CPU_IDX);
+    FaabricCpuSet& operator=(const FaabricCpuSet&) = delete;
+    FaabricCpuSet(const FaabricCpuSet&) = delete;
+
+    ~FaabricCpuSet();
+
+    cpu_set_t* get() { return &cpuSet; }
+
+  private:
+    cpu_set_t cpuSet;
+
+    // CPU index in internal CPU accounting
+    int cpuIdx = NO_CPU_IDX;
+};
+
+// Pin thread to any "unpinned" CPUs. Returns the CPU set it was pinned to.
+// We return a unique pointer to enforce RAII on the pinned-to CPU
+std::unique_ptr<FaabricCpuSet> pinThreadToFreeCpu(pthread_t thread);
+}