Merge pull request #455 from paulip1792/reserve-hugepages-by-numa-node

add option to reserve huge os pages at a specific numa node.

Author: daanx

doc/mimalloc-doc.h

@@ -782,6 +782,7 @@ typedef enum mi_option_e {
   mi_option_eager_region_commit, ///< Eagerly commit large (256MiB) memory regions (enabled by default, except on Windows)
   mi_option_large_os_pages,      ///< Use large OS pages (2MiB in size) if possible
   mi_option_reserve_huge_os_pages, ///< The number of huge OS pages (1GiB in size) to reserve at the start of the program.
+  mi_option_reserve_huge_os_pages_at, ///< Reserve huge OS pages at node N.
   mi_option_segment_cache,   ///< The number of segments per thread to keep cached.
   mi_option_page_reset,      ///< Reset page memory after \a mi_option_reset_delay milliseconds when it becomes free.
   mi_option_segment_reset,   ///< Experimental
@@ -1053,6 +1054,8 @@ or via environment variables.
    `MIMALLOC_EAGER_COMMIT_DELAY=N` (`N` is 1 by default) to delay the initial `N` segments (of 4MiB)
    of a thread to not allocate in the huge OS pages; this prevents threads that are short lived
    and allocate just a little to take up space in the huge OS page area (which cannot be reset).
+- `MIMALLOC_RESERVE_HUGE_OS_PAGES_AT=N`: where N is the numa node. This reserves the huge pages at a specific numa node. 
+   (`N` is -1 by default to reserve huge pages evenly among the given number of numa nodes (or use the available ones as detected))
 
 Use caution when using `fork` in combination with either large or huge OS pages: on a fork, the OS uses copy-on-write
 for all pages in the original process including the huge OS pages. When any memory is now written in that area, the

include/mimalloc.h

@@ -306,6 +306,7 @@ typedef enum mi_option_e {
   mi_option_reset_decommits,
   mi_option_large_os_pages,         // implies eager commit
   mi_option_reserve_huge_os_pages,
+  mi_option_reserve_huge_os_pages_at,
   mi_option_reserve_os_memory,
   mi_option_segment_cache,
   mi_option_page_reset,

readme.md

@@ -302,6 +302,8 @@ or via environment variables:
    `MIMALLOC_EAGER_COMMIT_DELAY=N` (`N` is 1 by default) to delay the initial `N` segments (of 4MiB)
    of a thread to not allocate in the huge OS pages; this prevents threads that are short lived
    and allocate just a little to take up space in the huge OS page area (which cannot be reset).
+- `MIMALLOC_RESERVE_HUGE_OS_PAGES_AT=N`: where N is the numa node. This reserves the huge pages at a specific numa node. 
+   (`N` is -1 by default to reserve huge pages evenly among the given number of numa nodes (or use the available ones as detected))
 
 Use caution when using `fork` in combination with either large or huge OS pages: on a fork, the OS uses copy-on-write
 for all pages in the original process including the huge OS pages. When any memory is now written in that area, the

src/init.c

@@ -494,7 +494,12 @@ void mi_process_init(void) mi_attr_noexcept {
 
   if (mi_option_is_enabled(mi_option_reserve_huge_os_pages)) {
     size_t pages = mi_option_get(mi_option_reserve_huge_os_pages);
-    mi_reserve_huge_os_pages_interleave(pages, 0, pages*500);
+    long reserve_at = mi_option_get(mi_option_reserve_huge_os_pages_at);
+    if (reserve_at != -1) {
+      mi_reserve_huge_os_pages_at(pages, reserve_at, pages*500);
+    } else {
+      mi_reserve_huge_os_pages_interleave(pages, 0, pages*500);
+    }
   } 
   if (mi_option_is_enabled(mi_option_reserve_os_memory)) {
     long ksize = mi_option_get(mi_option_reserve_os_memory);

src/options.c

@@ -76,6 +76,7 @@ static mi_option_desc_t options[_mi_option_last] =
   #endif
   { 0, UNINIT, MI_OPTION(large_os_pages) },      // use large OS pages, use only with eager commit to prevent fragmentation of VMA's
   { 0, UNINIT, MI_OPTION(reserve_huge_os_pages) },  // per 1GiB huge pages
+  { -1, UNINIT, MI_OPTION(reserve_huge_os_pages_at) }, // reserve huge pages at node N
   { 0, UNINIT, MI_OPTION(reserve_os_memory)     },
   { 0, UNINIT, MI_OPTION(segment_cache) },       // cache N segments per thread
   { 1, UNINIT, MI_OPTION(page_reset) },          // reset page memory on free