Print warning when too few logical cores are available

Author: facuMH

CMakeLists.txt

@@ -193,6 +193,12 @@ set(SOURCES
   "${CMAKE_CURRENT_BINARY_DIR}/src/version.cc"
 )
 
+if(WIN32)
+  set(SOURCES ${SOURCES} src/platform_specific/affinity.win.cc)
+ elseif(UNIX)
+  set(SOURCES ${SOURCES} src/platform_specific/affinity.unix.cc)
+endif()
+
 add_library(
   celerity_runtime
   STATIC

include/affinity.h

@@ -0,0 +1,15 @@
+#pragma once
+
+#include <cstdint>
+
+namespace celerity {
+namespace detail {
+
+	uint32_t affinity_cores_available();
+
+	/* a priori we need 3 threads, plus 1 for parallel-task workers and at least one more for host-task.
+	 This depends on the application invoking celerity. */
+	constexpr static uint64_t min_cores_needed = 5;
+
+} // namespace detail
+} // namespace celerity

include/utils.h

@@ -0,0 +1,23 @@
+#pragma once
+
+#include <cstdint>
+#include <type_traits>
+
+namespace celerity {
+namespace detail {
+	namespace utils {
+
+		template <typename BitMaskT>
+		constexpr inline uint32_t popcount(const BitMaskT bit_mask) noexcept {
+			static_assert(std::is_integral_v<BitMaskT> && std::is_unsigned_v<BitMaskT>, "popcount argument needs to be an unsigned integer type.");
+
+			uint32_t counter = 0;
+			for(auto b = bit_mask; b; b >>= 1) {
+				counter += b & 1;
+			}
+			return counter;
+		}
+
+	} // namespace utils
+} // namespace detail
+} // namespace celerity

src/platform_specific/affinity.unix.cc

@@ -0,0 +1,21 @@
+
+#include <cassert>
+#include <cstdint>
+
+#include <pthread.h>
+#include <sched.h>
+
+#include "affinity.h"
+
+namespace celerity {
+namespace detail {
+
+	uint32_t affinity_cores_available() {
+		cpu_set_t available_cores;
+		const auto affinity_error = pthread_getaffinity_np(pthread_self(), sizeof(cpu_set_t), &available_cores);
+		assert(affinity_error == 0 && "Error retrieving affinity mask.");
+		return CPU_COUNT(&available_cores);
+	}
+
+} // namespace detail
+} // namespace celerity

src/platform_specific/affinity.win.cc

@@ -0,0 +1,22 @@
+#include <cassert>
+
+#include <Windows.h>
+
+#include "affinity.h"
+#include "utils.h"
+
+namespace celerity {
+namespace detail {
+
+	uint32_t affinity_cores_available() {
+		using native_cpu_set = DWORD_PTR;
+
+		native_cpu_set available_cores;
+		[[maybe_unused]] native_cpu_set sys_affinity_mask;
+		const auto affinity_error = GetProcessAffinityMask(GetCurrentProcess(), &available_cores, &sys_affinity_mask);
+		assert(affinity_error != FALSE && "Error retrieving affinity mask.");
+		return utils::popcount(available_cores);
+	}
+
+} // namespace detail
+} // namespace celerity

src/runtime.cc

@@ -12,6 +12,7 @@
 
 #include <mpi.h>
 
+#include "affinity.h"
 #include "buffer.h"
 #include "buffer_manager.h"
 #include "command_graph.h"
@@ -23,6 +24,7 @@
 #include "scheduler.h"
 #include "task_manager.h"
 #include "user_bench.h"
+#include "utils.h"
 #include "version.h"
 
 namespace celerity {
@@ -98,8 +100,15 @@ namespace detail {
 		cfg = std::make_unique<config>(argc, argv, *default_logger);
 		graph_logger->set_level(cfg->get_log_level());
 
+		if(const uint32_t cores = affinity_cores_available(); cores < min_cores_needed) {
+			default_logger->warn("Celerity has detected that only {} logical cores are available to this process. It is recommended to assign at least {} "
+			                     "logical cores. Performance may be negatively impacted.",
+			    cores, min_cores_needed);
+		}
+
 		user_bench = std::make_unique<experimental::bench::detail::user_benchmarker>(*cfg, static_cast<node_id>(world_rank));
 
+
 		h_queue = std::make_unique<host_queue>(*default_logger);
 		d_queue = std::make_unique<device_queue>(*default_logger);
 

test/runtime_tests.cc

@@ -5,10 +5,17 @@
 #include <memory>
 #include <random>
 
+#ifdef _WIN32
+#include <Windows.h>
+#else
+#include <pthread.h>
+#endif
+
 #include <catch2/catch.hpp>
 
 #include <celerity.h>
 
+#include "affinity.h"
 #include "ranges.h"
 #include "region_map.h"
 
@@ -2305,5 +2312,25 @@ namespace detail {
 		}
 	}
 
+
+	TEST_CASE("affinity check", "[affinity]") {
+#ifdef _WIN32
+		SECTION("in Windows") {
+			DWORD_PTR cpu_mask = 1;
+			SetProcessAffinityMask(GetCurrentProcess(), cpu_mask);
+		}
+#else
+		SECTION("in Posix") {
+			cpu_set_t cpu_mask;
+			CPU_ZERO(&cpu_mask);
+			CPU_SET(0, &cpu_mask);
+			pthread_setaffinity_np(pthread_self(), sizeof(cpu_mask), &cpu_mask);
+		}
+#endif
+		const auto cores = affinity_cores_available();
+		REQUIRE(cores == 1);
+	}
+
+
 } // namespace detail
 } // namespace celerity