use multiple arenas in jemallocpool

fixed: #1076

Author: lplewa

include/umf/pools/pool_jemalloc.h

@@ -16,6 +16,33 @@ extern "C" {
 
 #include <umf/memory_pool_ops.h>
 
+struct umf_jemalloc_pool_params_t;
+
+/// @brief handle to the optional parameters of the jemalloc pool.
+typedef struct umf_jemalloc_pool_params_t *umf_jemalloc_pool_params_handle_t;
+
+/// @brief Create an optional struct to store parameters of jemalloc pool.
+/// @param hParams [out] handle to the newly created parameters struct.
+/// @return UMF_RESULT_SUCCESS on success or appropriate error code on failure.
+umf_result_t
+umfJemallocPoolParamsCreate(umf_jemalloc_pool_params_handle_t *hParams);
+
+/// @brief Destroy parameters struct.
+/// @param hParams handle to the parameters of the jemalloc pool.
+/// @return UMF_RESULT_SUCCESS on success or appropriate error code on failure.
+umf_result_t
+umfJemallocPoolParamsDestroy(umf_jemalloc_pool_params_handle_t hParams);
+
+/// @brief Customize number of arenas created for this pool. Default is the number of CPU cores * 4.
+/// \details
+/// The number of arenas is limited by jemalloc; setting this value too high may reduce the number of pools available for creation.
+/// @param hParams handle to the parameters of the jemalloc pool.
+/// @param numArenas number of arenas.
+/// @return UMF_RESULT_SUCCESS on success or appropriate error code on failure.
+umf_result_t
+umfJemallocPoolParamsSetNumArenas(umf_jemalloc_pool_params_handle_t hParams,
+                                  size_t numArenas);
+
 umf_memory_pool_ops_t *umfJemallocPoolOps(void);
 
 #ifdef __cplusplus

src/libumf.def

@@ -35,6 +35,9 @@ EXPORTS
     umfGetIPCHandle
     umfGetLastFailedMemoryProvider
     umfJemallocPoolOps
+    umfJemallocPoolParamsCreate
+    umfJemallocPoolParamsDestroy
+    umfJemallocPoolParamsSetNumArenas
     umfLevelZeroMemoryProviderOps
     umfLevelZeroMemoryProviderParamsCreate
     umfLevelZeroMemoryProviderParamsDestroy

src/libumf.map

@@ -29,6 +29,9 @@ UMF_0.10 {
         umfGetIPCHandle;
         umfGetLastFailedMemoryProvider;
         umfJemallocPoolOps;
+        umfJemallocPoolParamsCreate;
+        umfJemallocPoolParamsDestroy;
+        umfJemallocPoolParamsSetNumArenas;
         umfLevelZeroMemoryProviderOps;
         umfLevelZeroMemoryProviderParamsCreate;
         umfLevelZeroMemoryProviderParamsDestroy;

src/pool/pool_jemalloc.c

@@ -30,9 +30,14 @@ umf_memory_pool_ops_t *umfJemallocPoolOps(void) { return NULL; }
 
 #define MALLOCX_ARENA_MAX (MALLCTL_ARENAS_ALL - 1)
 
+typedef struct umf_jemalloc_pool_params_t {
+    size_t numArenas;
+} umf_jemalloc_pool_params_t;
+
 typedef struct jemalloc_memory_pool_t {
     umf_memory_provider_handle_t provider;
-    unsigned int arena_index; // index of jemalloc arena
+    size_t n_arenas;
+    unsigned int arena_index[];
 } jemalloc_memory_pool_t;
 
 static __TLS umf_result_t TLS_last_allocation_error;
@@ -47,6 +52,14 @@ static jemalloc_memory_pool_t *get_pool_by_arena_index(unsigned arena_ind) {
     return pool_by_arena_index[arena_ind];
 }
 
+// SplitMix64 hash
+static uint64_t hash64(uint64_t x) {
+    x += 0x9e3779b97f4a7c15;
+    x = (x ^ (x >> 30)) * 0xbf58476d1ce4e5b9;
+    x = (x ^ (x >> 27)) * 0x94d049bb133111eb;
+    return x ^ (x >> 31);
+}
+
 // arena_extent_alloc - an extent allocation function conforms to the extent_alloc_t type and upon
 // success returns a pointer to size bytes of mapped memory on behalf of arena arena_ind such that
 // the extent's base address is a multiple of alignment, as well as setting *zero to indicate
@@ -285,12 +298,17 @@ static extent_hooks_t arena_extent_hooks = {
     .merge = arena_extent_merge,
 };
 
+static unsigned get_arena_index(jemalloc_memory_pool_t *pool) {
+    unsigned pid = utils_getpid();
+    return pool->arena_index[hash64(pid) % pool->n_arenas];
+}
+
 static void *op_malloc(void *pool, size_t size) {
     assert(pool);
     jemalloc_memory_pool_t *je_pool = (jemalloc_memory_pool_t *)pool;
     // MALLOCX_TCACHE_NONE is set, because jemalloc can mix objects from different arenas inside
     // the tcache, so we wouldn't be able to guarantee isolation of different providers.
-    int flags = MALLOCX_ARENA(je_pool->arena_index) | MALLOCX_TCACHE_NONE;
+    int flags = MALLOCX_ARENA(get_arena_index(je_pool)) | MALLOCX_TCACHE_NONE;
     void *ptr = je_mallocx(size, flags);
     if (ptr == NULL) {
         TLS_last_allocation_error = UMF_RESULT_ERROR_OUT_OF_HOST_MEMORY;
@@ -343,7 +361,7 @@ static void *op_realloc(void *pool, void *ptr, size_t size) {
     jemalloc_memory_pool_t *je_pool = (jemalloc_memory_pool_t *)pool;
     // MALLOCX_TCACHE_NONE is set, because jemalloc can mix objects from different arenas inside
     // the tcache, so we wouldn't be able to guarantee isolation of different providers.
-    int flags = MALLOCX_ARENA(je_pool->arena_index) | MALLOCX_TCACHE_NONE;
+    int flags = MALLOCX_ARENA(get_arena_index(je_pool)) | MALLOCX_TCACHE_NONE;
     void *new_ptr = je_rallocx(ptr, size, flags);
     if (new_ptr == NULL) {
         TLS_last_allocation_error = UMF_RESULT_ERROR_OUT_OF_HOST_MEMORY;
@@ -364,7 +382,8 @@ static void *op_realloc(void *pool, void *ptr, size_t size) {
 static void *op_aligned_alloc(void *pool, size_t size, size_t alignment) {
     assert(pool);
     jemalloc_memory_pool_t *je_pool = (jemalloc_memory_pool_t *)pool;
-    unsigned arena = je_pool->arena_index;
+
+    unsigned arena = get_arena_index(je_pool);
     // MALLOCX_TCACHE_NONE is set, because jemalloc can mix objects from different arenas inside
     // the tcache, so we wouldn't be able to guarantee isolation of different providers.
     int flags =
@@ -382,62 +401,89 @@ static void *op_aligned_alloc(void *pool, size_t size, size_t alignment) {
 
 static umf_result_t op_initialize(umf_memory_provider_handle_t provider,
                                   void *params, void **out_pool) {
-    (void)params; // unused
     assert(provider);
     assert(out_pool);
 
     extent_hooks_t *pHooks = &arena_extent_hooks;
     size_t unsigned_size = sizeof(unsigned);
     int err;
+    umf_jemalloc_pool_params_t *jemalloc_params =
+        (umf_jemalloc_pool_params_t *)params;
 
-    jemalloc_memory_pool_t *pool =
-        umf_ba_global_alloc(sizeof(jemalloc_memory_pool_t));
-    if (!pool) {
-        return UMF_RESULT_ERROR_OUT_OF_HOST_MEMORY;
+    size_t nArenas = 0;
+    if (jemalloc_params) {
+        nArenas = jemalloc_params->numArenas;
     }
 
-    pool->provider = provider;
-
-    unsigned arena_index;
-    err = je_mallctl("arenas.create", (void *)&arena_index, &unsigned_size,
-                     NULL, 0);
-    if (err) {
-        LOG_ERR("Could not create arena.");
-        goto err_free_pool;
+    if (nArenas == 0) {
+        nArenas = utils_get_num_cores() * 4;
     }
 
-    // setup extent_hooks for newly created arena
-    char cmd[64];
-    snprintf(cmd, sizeof(cmd), "arena.%u.extent_hooks", arena_index);
-    err = je_mallctl(cmd, NULL, NULL, (void *)&pHooks, sizeof(void *));
-    if (err) {
-        snprintf(cmd, sizeof(cmd), "arena.%u.destroy", arena_index);
-        (void)je_mallctl(cmd, NULL, 0, NULL, 0);
-        LOG_ERR("Could not setup extent_hooks for newly created arena.");
-        goto err_free_pool;
+    jemalloc_memory_pool_t *pool = umf_ba_global_alloc(
+        sizeof(*pool) + nArenas * sizeof(*pool->arena_index));
+    if (!pool) {
+        return UMF_RESULT_ERROR_OUT_OF_HOST_MEMORY;
     }
 
-    pool->arena_index = arena_index;
-    pool_by_arena_index[arena_index] = pool;
-
+    pool->provider = provider;
+    pool->n_arenas = nArenas;
+
+    size_t num_created = 0;
+    for (size_t i = 0; i < nArenas; i++) {
+        unsigned arena_index;
+        err = je_mallctl("arenas.create", (void *)&arena_index, &unsigned_size,
+                         NULL, 0);
+        if (err) {
+            LOG_ERR("Could not create arena.");
+            goto err_cleanup;
+        }
+
+        pool->arena_index[num_created++] = arena_index;
+        if (arena_index >= MALLOCX_ARENA_MAX) {
+            LOG_ERR("Number of arenas exceeds the limit.");
+            goto err_cleanup;
+        }
+
+        pool_by_arena_index[arena_index] = pool;
+
+        // Setup extent_hooks for the newly created arena.
+        char cmd[64];
+        snprintf(cmd, sizeof(cmd), "arena.%u.extent_hooks", arena_index);
+        err = je_mallctl(cmd, NULL, NULL, (void *)&pHooks, sizeof(void *));
+        if (err) {
+            LOG_ERR("Could not setup extent_hooks for newly created arena.");
+            goto err_cleanup;
+        }
+    }
     *out_pool = (umf_memory_pool_handle_t)pool;
 
     VALGRIND_DO_CREATE_MEMPOOL(pool, 0, 0);
 
     return UMF_RESULT_SUCCESS;
 
-err_free_pool:
+err_cleanup:
+    // Destroy any arenas that were successfully created.
+    for (size_t i = 0; i < num_created; i++) {
+        char cmd[64];
+        unsigned arena = pool->arena_index[i];
+        snprintf(cmd, sizeof(cmd), "arena.%u.destroy", arena);
+        (void)je_mallctl(cmd, NULL, 0, NULL, 0);
+        pool_by_arena_index[arena] = NULL;
+    }
     umf_ba_global_free(pool);
     return UMF_RESULT_ERROR_MEMORY_PROVIDER_SPECIFIC;
 }
 
 static void op_finalize(void *pool) {
     assert(pool);
     jemalloc_memory_pool_t *je_pool = (jemalloc_memory_pool_t *)pool;
-    char cmd[64];
-    snprintf(cmd, sizeof(cmd), "arena.%u.destroy", je_pool->arena_index);
-    (void)je_mallctl(cmd, NULL, 0, NULL, 0);
-    pool_by_arena_index[je_pool->arena_index] = NULL;
+    for (size_t i = 0; i < je_pool->n_arenas; i++) {
+        char cmd[64];
+        unsigned arena = je_pool->arena_index[i];
+        snprintf(cmd, sizeof(cmd), "arena.%u.destroy", arena);
+        (void)je_mallctl(cmd, NULL, 0, NULL, 0);
+        pool_by_arena_index[arena] = NULL;
+    }
     umf_ba_global_free(je_pool);
 
     VALGRIND_DO_DESTROY_MEMPOOL(pool);
@@ -469,4 +515,32 @@ static umf_memory_pool_ops_t UMF_JEMALLOC_POOL_OPS = {
 umf_memory_pool_ops_t *umfJemallocPoolOps(void) {
     return &UMF_JEMALLOC_POOL_OPS;
 }
+
+umf_result_t
+umfJemallocPoolParamsCreate(umf_jemalloc_pool_params_handle_t *hParams) {
+    umf_jemalloc_pool_params_t *params = umf_ba_global_alloc(sizeof(*params));
+    if (!params) {
+        return UMF_RESULT_ERROR_OUT_OF_HOST_MEMORY;
+    }
+    *hParams = params;
+    return UMF_RESULT_SUCCESS;
+}
+
+umf_result_t
+umfJemallocPoolParamsDestroy(umf_jemalloc_pool_params_handle_t hParams) {
+    umf_ba_global_free(hParams);
+    return UMF_RESULT_SUCCESS;
+}
+
+umf_result_t
+umfJemallocPoolParamsSetNumArenas(umf_jemalloc_pool_params_handle_t hParams,
+                                  size_t numArenas) {
+    if (!hParams) {
+        LOG_ERR("jemalloc pool params handle is NULL");
+        return UMF_RESULT_ERROR_INVALID_ARGUMENT;
+    }
+    hParams->numArenas = numArenas;
+    return UMF_RESULT_SUCCESS;
+}
+
 #endif /* UMF_POOL_JEMALLOC_ENABLED */

src/utils/utils_common.h

@@ -115,6 +115,9 @@ int utils_getpid(void);
 // get the current thread ID
 int utils_gettid(void);
 
+// get the number of CPU cores
+unsigned utils_get_num_cores(void);
+
 // close file descriptor
 int utils_close_fd(int fd);
 

src/utils/utils_posix_common.c

@@ -37,13 +37,17 @@
          ? LLONG_MAX                                                           \
          : (sizeof(off_t) == sizeof(long) ? LONG_MAX : INT_MAX))
 
-static UTIL_ONCE_FLAG Page_size_is_initialized = UTIL_ONCE_FLAG_INIT;
+static UTIL_ONCE_FLAG System_info_is_initialized = UTIL_ONCE_FLAG_INIT;
 static size_t Page_size;
+static unsigned Core_conut;
 
-static void _utils_get_page_size(void) { Page_size = sysconf(_SC_PAGE_SIZE); }
+static void _utils_get_system_info(void) {
+    Page_size = sysconf(_SC_PAGE_SIZE);
+    Core_conut = sysconf(_SC_NPROCESSORS_ONLN);
+}
 
 size_t utils_get_page_size(void) {
-    utils_init_once(&Page_size_is_initialized, _utils_get_page_size);
+    utils_init_once(&System_info_is_initialized, _utils_get_system_info);
     return Page_size;
 }
 
@@ -62,6 +66,11 @@ int utils_gettid(void) {
 #endif
 }
 
+unsigned utils_get_num_cores(void) {
+    utils_init_once(&System_info_is_initialized, _utils_get_system_info);
+    return Core_conut;
+}
+
 int utils_close_fd(int fd) { return close(fd); }
 
 umf_result_t utils_errno_to_umf_result(int err) {

src/utils/utils_windows_common.c

@@ -1,6 +1,6 @@
 /*
  *
- * Copyright (C) 2024 Intel Corporation
+ * Copyright (C) 2024-2025 Intel Corporation
  *
  * Under the Apache License v2.0 with LLVM Exceptions. See LICENSE.TXT.
  * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
@@ -22,20 +22,27 @@
 
 #define BUFFER_SIZE 1024
 
-static UTIL_ONCE_FLAG Page_size_is_initialized = UTIL_ONCE_FLAG_INIT;
+static UTIL_ONCE_FLAG System_info_is_initialized = UTIL_ONCE_FLAG_INIT;
 static size_t Page_size;
+static unsigned Core_count;
 
-static void _utils_get_page_size(void) {
+static void _utils_get_system_info(void) {
     SYSTEM_INFO SystemInfo;
     GetSystemInfo(&SystemInfo);
     Page_size = SystemInfo.dwPageSize;
+    Core_count = SystemInfo.dwNumberOfProcessors;
 }
 
 size_t utils_get_page_size(void) {
-    utils_init_once(&Page_size_is_initialized, _utils_get_page_size);
+    utils_init_once(&System_info_is_initialized, _utils_get_system_info);
     return Page_size;
 }
 
+unsigned utils_get_num_cores(void) {
+    utils_init_once(&System_info_is_initialized, _utils_get_system_info);
+    return Core_count;
+}
+
 int utils_getpid(void) { return GetCurrentProcessId(); }
 
 int utils_gettid(void) { return GetCurrentThreadId(); }