hbwmalloc - Man Page

hbwmalloc — The high bandwidth memory interface
Note: hbwmalloc.h functionality is considered as stable API (STANDARD API).

#include <hbwmalloc.h>

Link with -lmemkind

int hbw_check_available(void);
void* hbw_malloc(size_t size);
void* hbw_calloc(size_t nmemb, size_t size);
void* hbw_realloc (void *ptr, size_t size);
void hbw_free(void *ptr);
size_t hbw_malloc_usable_size(void *ptr);
int hbw_posix_memalign(void **memptr, size_t alignment, size_t size);
int hbw_posix_memalign_psize(void **memptr, size_t alignment, size_t size, hbw_pagesize_t pagesize);
hbw_policy_t hbw_get_policy(void);
int hbw_set_policy(hbw_policy_t mode);
int hbw_verify_memory_region(void *addr, size_t size, int flags);

hbw_check_available() returns zero if high bandwidth memory is available or an error code described in the Errors section if not.

hbw_malloc() allocates size bytes of uninitialized high bandwidth memory. The allocated space is suitably aligned (after possible pointer coercion) for storage of any type of object. If size is zero then hbw_malloc() returns NULL.

hbw_calloc() allocates space for nmemb objects in high bandwidth memory, each size bytes in length. The result is identical to calling hbw_malloc() with an argument of nmemb * size, with the exception that the allocated memory is explicitly initialized to zero bytes.  If nmemb or size is 0, then hbw_calloc() returns NULL.

hbw_realloc() changes the size of the previously allocated high bandwidth memory referenced by ptr to size bytes. The contents of the memory remain unchanged up to the lesser of the new and old sizes. If the new size is larger, the contents of the newly allocated portion of the memory are undefined. Upon success, the memory referenced by ptr is freed and a pointer to the newly allocated high bandwidth memory is returned.

Note: hbw_realloc() may move the memory allocation, resulting in a different return value than ptr.

If ptr is NULL, the hbw_realloc() function behaves identically to hbw_malloc() for the specified size. If size is equal to zero, and ptr is not NULL, then the call is equivalent to hbw_free(ptr) and NULL is returned. The address ptr, if not NULL, was returned by a previous call to hbw_malloc(), hbw_calloc(), hbw_realloc() or hbw_posix_memalign(). Otherwise, or if hbw_free(ptr) was called before, undefined behavior occurs.

Note: hbw_realloc() cannot be used with a pointer returned by hbw_posix_memalign_psize().

hbw_free() causes the allocated memory referenced by ptr to be made available for future allocations. If ptr is NULL, no action occurs. The address ptr, if not NULL, must have been returned by a previous call to hbw_malloc(), hbw_calloc(), hbw_realloc(), hbw_posix_memalign() or hbw_posix_memalign_psize(). Otherwise, if hbw_free(ptr) was called before, undefined behavior occurs.

hbw_malloc_usable_size() returns the number of usable bytes in the block pointed to by ptr, a pointer to a block of memory allocated by hbw_malloc(), hbw_calloc(), hbw_realloc(), hbw_posix_memalign(), or hbw_posix_memalign_psize().

hbw_posix_memalign() allocates size bytes of high bandwidth memory such that the allocation's base address is an even multiple of alignment, and returns the allocation in the value pointed to by memptr. The requested alignment must be a power of 2 at least as large as sizeof(void*). If size is 0, then hbw_posix_memalign() returns 0, with a NULL returned in memptr.

hbw_posix_memalign_psize() allocates size bytes of high bandwidth memory such that the allocation's base address is an even multiple of alignment, and returns the allocation in the value pointed to by memptr. The requested alignment must be a power of 2 at least as large as sizeof(void*). The memory will be allocated using pages determined by the pagesize variable which may be one of the following enumerated values:

HBW_PAGESIZE_4KB

The four kilobyte page size option. Note that with transparent huge pages enabled these allocations may be promoted by the operating system to two megabyte pages.

HBW_PAGESIZE_2MB

The two megabyte page size option. Note: This page size requires huge pages configuration described in System Configuration section.

HBW_PAGESIZE_1GB (DEPRECATED)

This option allows the user to specify arbitrary sizes backed by 1GB chunks of huge pages. Huge pages are allocated even if the size is not a modulo of 1GB. Note: This page size requires huge pages configuration described in System Configuration section.

HBW_PAGESIZE_1GB_STRICT (DEPRECATED)

The total size of the allocation must be a multiple of 1GB with this option, otherwise the allocation will fail. Note: This page size requires huge pages configuration described in System Configuration section.

Note: HBW_PAGESIZE_2MB, HBW_PAGESIZE_1GB and HBW_PAGESIZE_1GB_STRICT options are not supported with HBW_POLICY_INTERLEAVE policy.

hbw_get_policy() returns the current fallback policy when insufficient high bandwidth memory is available.

hbw_set_policy() sets the current fallback policy. The policy can be modified only once in the lifetime of an application and before calling hbw_malloc(), hbw_calloc(), hbw_realloc(), hbw_posix_memalign(), or hbw_posix_memalign_psize() function.
Note: If the policy is not set, than HBW_POLICY_PREFERRED will be used by default.

HBW_POLICY_BIND

If insufficient high bandwidth memory from the nearest NUMA node is available to satisfy a request, the allocated pointer is set to NULL and errno is set to ENOMEM. If insufficient high bandwidth memory pages are available at fault time the Out Of Memory (OOM) Killer is triggered. Note that pages are faulted exclusively from the high bandwidth NUMA node nearest at time of allocation, not at time of fault.

HBW_POLICY_BIND_ALL

If insufficient high bandwidth memory is available to satisfy a request, the allocated pointer is set to NULL and errno is set to ENOMEM. If insufficient high bandwidth memory pages are available at fault time the Out Of Memory (OOM) Killer is triggered. Note that pages are faulted from the high bandwidth NUMA nodes. Nearest NUMA node is selected at time of page fault.

HBW_POLICY_PREFERRED

If insufficient memory is available from the high bandwidth NUMA node closest at allocation time, fall back to standard memory (default) with the smallest NUMA distance.

HBW_POLICY_INTERLEAVE

Interleave faulted pages from across all high bandwidth NUMA nodes using standard size pages (the Transparent Huge Page feature is disabled).

hbw_verify_memory_region() verifies if memory region fully falls into high bandwidth memory. Returns 0 if memory address range from addr to addr + size is allocated in high bandwidth memory, -1 if any fragment of memory was not backed by high bandwidth memory (e.g. when memory is not initialized) or one of error codes described in Errors section.

Using this function in production code may result in serious performance penalty.

The Flags argument may include optional flags that modify function behavior:

HBW_TOUCH_PAGES

Before checking pages, function will touch first byte of all pages in address range starting from addr to addr + size by read and write (so the content will be overwritten by the same data as it was read). Using this option may trigger Out Of Memory Killer.

hbw_get_policy() returns HBW_POLICY_BIND, HBW_POLICY_BIND_ALL, HBW_POLICY_PREFERRED or HBW_POLICY_INTERLEAVE which represents the current high bandwidth policy. hbw_free() do not have return value. hbw_malloc() hbw_calloc() and hbw_realloc() return the pointer to the allocated memory, or NULL if the request fails. hbw_posix_memalign(), hbw_posix_memalign_psize() and hbw_set_policy() return zero on success and return an error code as described in the Errors section below on failure.

Error codes described here are the POSIX standard error codes as defined in

<errno.h>

hbw_check_available()

returns ENODEV if high-bandwidth memory is unavailable.

hbw_posix_memalign() and hbw_posix_memalign_psize()

If the alignment parameter is not a power of two, or was not a multiple of sizeof(void*), then EINVAL is returned. If the policy and pagesize combination is unsupported then EINVAL is returned. If there was insufficient memory to satisfy the request then ENOMEM is returned.

hbw_set_policy()

returns EPERM if hbw_set_policy() was called more than once, or EINVAL if mode argument was neither HBW_POLICY_PREFERRED, HBW_POLICY_BIND, HBW_POLICY_BIND_ALL nor HBW_POLICY_INTERLEAVE.

hbw_verify_memory_region()

returns EINVAL if addr is NULL, size equals 0 or flags contained unsupported bit set. If memory pointed by addr could not be verified then EFAULT is returned.

The <hbwmalloc.h> file defines the external functions and enumerations for the hbwmalloc library. These interfaces define a heap manager that targets high bandwidth memory numa nodes.

/usr/bin/memkind-hbw-nodes

Prints a comma-separated list of high bandwidth nodes.

MEMKIND_HBW_NODES

This environment variable is a comma-separated list of NUMA nodes that are treated as high bandwidth. Uses the libnuma routine numa_parse_nodestring() for parsing, so the syntax described in the numa(3) man page for this routine applies for example: 1-3,5 is a valid setting.

MEMKIND_ARENA_NUM_PER_KIND

This environment variable allows leveraging internal mechanism of the library for setting number of arenas per kind. Value should be a positive integer (not greater than INT_MAX defined in <limits.h>). The user should set the value based on the characteristics of application that is using the library. Higher value can provide better performance in extremely multithreaded applications at the cost of memory overhead. See section IMPLEMENTATION Notes of jemalloc(3) for more details about arenas.

MEMKIND_HEAP_MANAGER

Controls heap management behavior in memkind library by switching to one of the available heap managers.
Values:
   JEMALLOC - sets the jemalloc heap manager
   TBB - sets the Intel Threading Building Blocks heap manager. This option requires installed
   Intel Threading Building Blocks library.

Note: If the MEMKIND_HEAP_MANAGER is not set then the jemalloc heap manager will be used by default.

Interfaces for obtaining 2MB (HUGETLB) memory need allocated huge pages in the kernel's huge page pool.

HUGETLB (huge pages)

Current number of "persistent" huge pages can be read from /proc/sys/vm/nr_hugepages file. Proposed way of setting hugepages is: sudo sysctl vm.nr_hugepages=<number_of_hugepages>. More information can be found here: https://www.kernel.org/doc/Documentation/vm/hugetlbpage.txt

HUGETLB (huge pages)

There might be some overhead in huge pages consumption caused by heap management. If your allocation fails because of OOM, please try to allocate extra huge pages (e.g. 8 huge pages).

Copyright (C) 2014 - 2020 Intel Corporation. All rights reserved.

malloc(3), numa(3), numactl(8), mbind(2), mmap(2), move_pages(2), jemalloc(3), memkind(3)

hbwallocator(3), memkind-hbw-nodes(1).