Merge pull request #408 from kswiecicki/memspace-hbw

Add memspace "highest bandwidth"

Author: lplewa

README.md

@@ -223,6 +223,11 @@ using umfMemspaceHostAllGet.
 Memspace backed by all available NUMA nodes discovered on the platform sorted by capacity.
 Can be retrieved using umfMemspaceHighestCapacityGet.
 
+#### Highest bandwidth memspace
+
+Memspace backed by an aggregated list of NUMA nodes identified as highest bandwidth after selecting each available NUMA node as the initiator.
+Querying the bandwidth value requires HMAT support on the platform. Calling `umfMemspaceHighestBandwidthGet()` will return NULL if it's not supported.
+
 ### Proxy library
 
 UMF provides the UMF proxy library (`umf_proxy`) that makes it possible

include/umf/memspace.h

@@ -56,6 +56,11 @@ umf_memspace_handle_t umfMemspaceHostAllGet(void);
 ///
 umf_memspace_handle_t umfMemspaceHighestCapacityGet(void);
 
+/// \brief Retrieves predefined highest bandwidth memspace.
+/// \return highest bandwidth memspace handle on success or NULL on failure.
+///
+umf_memspace_handle_t umfMemspaceHighestBandwidthGet(void);
+
 #ifdef __cplusplus
 }
 #endif

scripts/qemu/run-build.sh

@@ -39,7 +39,7 @@ make -j $(nproc)
 echo password | sudo sync;
 echo password | sudo sh -c "/usr/bin/echo 3 > /proc/sys/vm/drop_caches"
 
-ctest --output-on-failure
+ctest --verbose
 
 # run tests bound to a numa node
 numactl -N 0 ctest --output-on-failure

src/CMakeLists.txt

@@ -82,7 +82,8 @@ set(UMF_SOURCES_COMMON_LINUX_MACOSX
     memory_targets/memory_target_numa.c
     memspaces/memspace_numa.c
     memspaces/memspace_host_all.c
-    memspaces/memspace_highest_capacity.c)
+    memspaces/memspace_highest_capacity.c
+    memspaces/memspace_highest_bandwidth.c)
 
 set(UMF_SOURCES_LINUX ${UMF_SOURCES_LINUX} ${UMF_SOURCES_COMMON_LINUX_MACOSX}
                       provider/provider_os_memory_linux.c)

src/libumf.map

@@ -30,6 +30,7 @@ UMF_1.0 {
         umfMemoryProviderPutIPCHandle;
         umfMemspaceCreateFromNumaArray;
         umfMemspaceDestroy;
+        umfMemspaceHighestBandwidthGet;
         umfMemspaceHighestCapacityGet;
         umfMemspaceHostAllGet;
         umfOpenIPCHandle;

src/libumf_linux.c

@@ -29,6 +29,7 @@ void __attribute__((destructor)) umfDestroy(void) {
     umfMemoryTrackerDestroy(t);
     umfMemspaceHostAllDestroy();
     umfMemspaceHighestCapacityDestroy();
+    umfMemspaceHighestBandwidthDestroy();
     umfDestroyTopology();
 }
 

src/memory_target.c

@@ -79,7 +79,21 @@ umf_result_t umfMemoryTargetClone(umf_memory_target_handle_t memoryTarget,
 
 umf_result_t umfMemoryTargetGetCapacity(umf_memory_target_handle_t memoryTarget,
                                         size_t *capacity) {
-    assert(memoryTarget);
-    assert(capacity);
+    if (!memoryTarget || !capacity) {
+        return UMF_RESULT_ERROR_INVALID_ARGUMENT;
+    }
+
     return memoryTarget->ops->get_capacity(memoryTarget->priv, capacity);
 }
+
+umf_result_t
+umfMemoryTargetGetBandwidth(umf_memory_target_handle_t srcMemoryTarget,
+                            umf_memory_target_handle_t dstMemoryTarget,
+                            size_t *bandwidth) {
+    if (!srcMemoryTarget || !dstMemoryTarget || !bandwidth) {
+        return UMF_RESULT_ERROR_INVALID_ARGUMENT;
+    }
+
+    return srcMemoryTarget->ops->get_bandwidth(
+        srcMemoryTarget->priv, dstMemoryTarget->priv, bandwidth);
+}

src/memory_target.h

@@ -37,6 +37,10 @@ umf_result_t umfMemoryTargetClone(umf_memory_target_handle_t memoryTarget,
                                   umf_memory_target_handle_t *outHandle);
 umf_result_t umfMemoryTargetGetCapacity(umf_memory_target_handle_t memoryTarget,
                                         size_t *capacity);
+umf_result_t
+umfMemoryTargetGetBandwidth(umf_memory_target_handle_t srcMemoryTarget,
+                            umf_memory_target_handle_t dstMemoryTarget,
+                            size_t *bandwidth);
 
 #ifdef __cplusplus
 }

src/memory_target_ops.h

@@ -41,6 +41,8 @@ typedef struct umf_memory_target_ops_t {
         umf_memory_provider_handle_t *provider);
 
     umf_result_t (*get_capacity)(void *memoryTarget, size_t *capacity);
+    umf_result_t (*get_bandwidth)(void *srcMemoryTarget, void *dstMemoryTarget,
+                                  size_t *bandwidth);
 } umf_memory_target_ops_t;
 
 #ifdef __cplusplus

src/memory_targets/memory_target_numa.c

@@ -19,6 +19,7 @@
 #include "base_alloc_global.h"
 #include "memory_target_numa.h"
 #include "topology.h"
+#include "utils_log.h"
 
 struct numa_memory_target_t {
     unsigned physical_id;
@@ -124,6 +125,10 @@ static umf_result_t numa_clone(void *memTarget, void **outMemTarget) {
 }
 
 static umf_result_t numa_get_capacity(void *memTarget, size_t *capacity) {
+    if (!memTarget || !capacity) {
+        return UMF_RESULT_ERROR_INVALID_ARGUMENT;
+    }
+
     hwloc_topology_t topology = umfGetTopology();
     if (!topology) {
         return UMF_RESULT_ERROR_NOT_SUPPORTED;
@@ -143,12 +148,61 @@ static umf_result_t numa_get_capacity(void *memTarget, size_t *capacity) {
     return UMF_RESULT_SUCCESS;
 }
 
+static umf_result_t numa_get_bandwidth(void *srcMemoryTarget,
+                                       void *dstMemoryTarget,
+                                       size_t *bandwidth) {
+    if (!srcMemoryTarget || !dstMemoryTarget || !bandwidth) {
+        return UMF_RESULT_ERROR_INVALID_ARGUMENT;
+    }
+
+    hwloc_topology_t topology = umfGetTopology();
+    if (!topology) {
+        return UMF_RESULT_ERROR_NOT_SUPPORTED;
+    }
+
+    hwloc_obj_t srcNumaNode = hwloc_get_obj_by_type(
+        topology, HWLOC_OBJ_NUMANODE,
+        ((struct numa_memory_target_t *)srcMemoryTarget)->physical_id);
+    if (!srcNumaNode) {
+        return UMF_RESULT_ERROR_INVALID_ARGUMENT;
+    }
+
+    hwloc_obj_t dstNumaNode = hwloc_get_obj_by_type(
+        topology, HWLOC_OBJ_NUMANODE,
+        ((struct numa_memory_target_t *)dstMemoryTarget)->physical_id);
+    if (!dstNumaNode) {
+        return UMF_RESULT_ERROR_INVALID_ARGUMENT;
+    }
+
+    // Given NUMA nodes aren't local, HWLOC returns an error in such case.
+    if (!hwloc_bitmap_intersects(srcNumaNode->cpuset, dstNumaNode->cpuset)) {
+        *bandwidth = 0;
+        return UMF_RESULT_SUCCESS;
+    }
+
+    struct hwloc_location initiator = {.location.cpuset = srcNumaNode->cpuset,
+                                       .type = HWLOC_LOCATION_TYPE_CPUSET};
+    hwloc_uint64_t value = 0;
+    int ret = hwloc_memattr_get_value(topology, HWLOC_MEMATTR_ID_BANDWIDTH,
+                                      dstNumaNode, &initiator, 0, &value);
+    if (ret) {
+        LOG_ERR("Retrieving bandwidth for initiator node %u to node %u failed.",
+                srcNumaNode->os_index, dstNumaNode->os_index);
+        return (errno == EINVAL) ? UMF_RESULT_ERROR_NOT_SUPPORTED
+                                 : UMF_RESULT_ERROR_UNKNOWN;
+    }
+
+    *bandwidth = value;
+    return UMF_RESULT_SUCCESS;
+}
+
 struct umf_memory_target_ops_t UMF_MEMORY_TARGET_NUMA_OPS = {
     .version = UMF_VERSION_CURRENT,
     .initialize = numa_initialize,
     .finalize = numa_finalize,
     .pool_create_from_memspace = numa_pool_create_from_memspace,
     .clone = numa_clone,
     .get_capacity = numa_get_capacity,
+    .get_bandwidth = numa_get_bandwidth,
     .memory_provider_create_from_memspace =
         numa_memory_provider_create_from_memspace};

src/memspace.c

@@ -208,3 +208,85 @@ umfMemspaceSortDesc(umf_memspace_handle_t hMemspace,
 
     return UMF_RESULT_SUCCESS;
 }
+
+umf_result_t umfMemspaceFilter(umf_memspace_handle_t hMemspace,
+                               umfGetTargetFn getTarget,
+                               umf_memspace_handle_t *filteredMemspace) {
+    if (!hMemspace || !getTarget) {
+        return UMF_RESULT_ERROR_INVALID_ARGUMENT;
+    }
+
+    umf_memory_target_handle_t *uniqueBestNodes =
+        umf_ba_global_alloc(hMemspace->size * sizeof(*uniqueBestNodes));
+    if (!uniqueBestNodes) {
+        return UMF_RESULT_ERROR_OUT_OF_HOST_MEMORY;
+    }
+
+    umf_result_t ret = UMF_RESULT_SUCCESS;
+
+    size_t numUniqueBestNodes = 0;
+    for (size_t nodeIdx = 0; nodeIdx < hMemspace->size; nodeIdx++) {
+        umf_memory_target_handle_t target = NULL;
+        ret = getTarget(hMemspace->nodes[nodeIdx], hMemspace->nodes,
+                        hMemspace->size, &target);
+        if (ret != UMF_RESULT_SUCCESS) {
+            goto err_free_best_targets;
+        }
+
+        // check if the target is already present in the best nodes
+        size_t bestTargetIdx;
+        for (bestTargetIdx = 0; bestTargetIdx < numUniqueBestNodes;
+             bestTargetIdx++) {
+            if (uniqueBestNodes[bestTargetIdx] == target) {
+                break;
+            }
+        }
+
+        // if the target is not present, add it to the best nodes
+        if (bestTargetIdx == numUniqueBestNodes) {
+            uniqueBestNodes[numUniqueBestNodes++] = target;
+        }
+    }
+
+    // copy the unique best nodes into a new memspace
+    umf_memspace_handle_t newMemspace =
+        umf_ba_global_alloc(sizeof(*newMemspace));
+    if (!newMemspace) {
+        ret = UMF_RESULT_ERROR_OUT_OF_HOST_MEMORY;
+        goto err_free_best_targets;
+    }
+
+    newMemspace->size = numUniqueBestNodes;
+    newMemspace->nodes =
+        umf_ba_global_alloc(sizeof(*newMemspace->nodes) * newMemspace->size);
+    if (!newMemspace->nodes) {
+        ret = UMF_RESULT_ERROR_OUT_OF_HOST_MEMORY;
+        goto err_free_new_memspace;
+    }
+
+    size_t cloneIdx = 0;
+    for (size_t cloneIdx = 0; cloneIdx < newMemspace->size; cloneIdx++) {
+        ret = umfMemoryTargetClone(uniqueBestNodes[cloneIdx],
+                                   &newMemspace->nodes[cloneIdx]);
+        if (ret != UMF_RESULT_SUCCESS) {
+            goto err_free_cloned_nodes;
+        }
+    }
+
+    *filteredMemspace = newMemspace;
+    umf_ba_global_free(uniqueBestNodes);
+
+    return UMF_RESULT_SUCCESS;
+
+err_free_cloned_nodes:
+    while (cloneIdx != 0) {
+        cloneIdx--;
+        umfMemoryTargetDestroy(newMemspace->nodes[cloneIdx]);
+    }
+    umf_ba_global_free(newMemspace->nodes);
+err_free_new_memspace:
+    umf_ba_global_free(newMemspace);
+err_free_best_targets:
+    umf_ba_global_free(uniqueBestNodes);
+    return ret;
+}

src/memspace_internal.h

@@ -39,6 +39,18 @@ typedef umf_result_t (*umfGetPropertyFn)(umf_memory_target_handle_t,
 umf_result_t umfMemspaceSortDesc(umf_memspace_handle_t hMemspace,
                                  umfGetPropertyFn getProperty);
 
+typedef umf_result_t (*umfGetTargetFn)(umf_memory_target_handle_t initiator,
+                                       umf_memory_target_handle_t *nodes,
+                                       size_t numNodes,
+                                       umf_memory_target_handle_t *target);
+
+///
+/// \brief Filters the targets using getTarget() to create a new memspace
+///
+umf_result_t umfMemspaceFilter(umf_memspace_handle_t hMemspace,
+                               umfGetTargetFn getTarget,
+                               umf_memspace_handle_t *filteredMemspace);
+
 ///
 /// \brief Destroys memspace
 /// \param hMemspace handle to memspace
@@ -47,6 +59,7 @@ void umfMemspaceDestroy(umf_memspace_handle_t hMemspace);
 
 void umfMemspaceHostAllDestroy(void);
 void umfMemspaceHighestCapacityDestroy(void);
+void umfMemspaceHighestBandwidthDestroy(void);
 
 #ifdef __cplusplus
 }