This patch adds the --obey-numa-mempol which requests that the
machine allocate hugepages based on the NUMA memory policy if
it is available.
Signed-off-by: Eric B Munson <ebmun...@us.ibm.com>
---
Changes from V1:
* Rename option to obey-mempolicy
* Remove if/else's around pool adjustment and replace with inline conditional
* Rework man page entry to more accurately discribe the new switch
---
hugeadm.c | 32 +++++++++++++++++++++++++++-----
hugeutils.c | 4 ++++
libhugetlbfs_privutils.h | 13 +++++++------
man/hugeadm.8 | 8 ++++++++
4 files changed, 46 insertions(+), 11 deletions(-)
diff --git a/hugeadm.c b/hugeadm.c
index ebc5bfe..5800cfe 100644
--- a/hugeadm.c
+++ b/hugeadm.c
@@ -91,6 +91,8 @@ void print_usage()
CONT("specified count on failure");
OPTION("--pool-pages-min
<size|DEFAULT>:[+|-]<pagecount|memsize<G|M|K>>", "");
CONT("Adjust pool 'size' lower bound");
+ OPTION("--obey-mempolicy", "Obey the NUMA memory policy when");
+ CONT("adjusting the pool 'size' lower bound");
OPTION("--pool-pages-max
<size|DEFAULT>:[+|-]<pagecount|memsize<G|M|K>>", "");
CONT("Adjust pool 'size' upper bound");
OPTION("--set-recommended-min_free_kbytes", "");
@@ -150,6 +152,7 @@ int opt_set_hugetlb_shm_group = 0;
int opt_temp_swap = 0;
int opt_ramdisk_swap = 0;
int opt_swap_persist = 0;
+int opt_obey_mempolicy = 0;
unsigned long opt_limit_mount_size = 0;
int opt_limit_mount_inodes = 0;
int verbose_level = VERBOSITY_DEFAULT;
@@ -231,6 +234,7 @@ void verbose_expose(void)
#define LONG_POOL_LIST (LONG_POOL|'l')
#define LONG_POOL_MIN_ADJ (LONG_POOL|'m')
#define LONG_POOL_MAX_ADJ (LONG_POOL|'M')
+#define LONG_POOL_MEMPOL (LONG_POOL|'p')
#define LONG_SET_RECOMMENDED_MINFREEKBYTES ('k' << 8)
#define LONG_SET_RECOMMENDED_SHMMAX ('x' << 8)
@@ -1176,8 +1180,18 @@ void pool_adjust(char *cmd, unsigned int counter)
add_ramdisk_swap(page_size);
check_swap();
}
- INFO("setting HUGEPAGES_TOTAL to %ld\n", min);
- set_huge_page_counter(page_size, HUGEPAGES_TOTAL, min);
+
+ if (opt_obey_mempolicy && get_huge_page_counter(page_size,
+ HUGEPAGES_TOTAL_MEMPOL) < 0) {
+ opt_obey_mempolicy = 0;
+ WARNING("Counter for NUMA huge page allocations is not found,
continuing with normal pool adjustment\n");
+ }
+
+ INFO("setting HUGEPAGES_TOTAL%s to %ld\n",
+ opt_obey_mempolicy ? "_MEMPOL" : "", min);
+ set_huge_page_counter(page_size,
+ opt_obey_mempolicy ? HUGEPAGES_TOTAL_MEMPOL : HUGEPAGES_TOTAL,
+ min);
get_pool_size(page_size, &pools[pos]);
/* If we fail to make an allocation, retry if user requests */
@@ -1191,9 +1205,12 @@ void pool_adjust(char *cmd, unsigned int counter)
sleep(6);
last_pool_value = pools[pos].minimum;
- INFO("Retrying allocation HUGEPAGES_TOTAL to %ld current %ld\n",
- min,
pools[pos].minimum);
- set_huge_page_counter(page_size, HUGEPAGES_TOTAL, min);
+ INFO("Retrying allocation HUGEPAGES_TOTAL%s to %ld current
%ld\n", opt_obey_mempolicy ? "_MEMPOL" : "", min, pools[pos].minimum);
+ set_huge_page_counter(page_size,
+ opt_obey_mempolicy ?
+ HUGEPAGES_TOTAL_MEMPOL :
+ HUGEPAGES_TOTAL,
+ min);
get_pool_size(page_size, &pools[pos]);
}
@@ -1280,6 +1297,7 @@ int main(int argc, char** argv)
{"pool-list", no_argument, NULL, LONG_POOL_LIST},
{"pool-pages-min", required_argument, NULL, LONG_POOL_MIN_ADJ},
{"pool-pages-max", required_argument, NULL, LONG_POOL_MAX_ADJ},
+ {"obey-mempolicy", no_argument, NULL, LONG_POOL_MEMPOL},
{"set-recommended-min_free_kbytes", no_argument, NULL,
LONG_SET_RECOMMENDED_MINFREEKBYTES},
{"set-recommended-shmmax", no_argument, NULL,
LONG_SET_RECOMMENDED_SHMMAX},
{"set-shm-group", required_argument, NULL,
LONG_SET_HUGETLB_SHM_GROUP},
@@ -1379,6 +1397,10 @@ int main(int argc, char** argv)
}
break;
+ case LONG_POOL_MEMPOL:
+ opt_obey_mempolicy = 1;
+ break;
+
case LONG_POOL_MAX_ADJ:
if (! kernel_has_overcommit()) {
ERROR("kernel does not support overcommit, "
diff --git a/hugeutils.c b/hugeutils.c
index 1e35597..3a9501f 100644
--- a/hugeutils.c
+++ b/hugeutils.c
@@ -146,6 +146,10 @@ static struct hugetlb_pool_counter_info_t
hugetlb_counter_info[] = {
.meminfo_key = "HugePages_Total:",
.sysfs_file = "nr_hugepages",
},
+ [HUGEPAGES_TOTAL_MEMPOL] = {
+ .meminfo_key = "HugePages_Total:",
+ .sysfs_file = "nr_hugepages_mempolicy",
+ },
[HUGEPAGES_FREE] = {
.meminfo_key = "HugePages_Free:",
.sysfs_file = "free_hugepages",
diff --git a/libhugetlbfs_privutils.h b/libhugetlbfs_privutils.h
index 18bcedb..aaa638f 100644
--- a/libhugetlbfs_privutils.h
+++ b/libhugetlbfs_privutils.h
@@ -34,12 +34,13 @@
/* Hugetlb pool counter operations */
/* Keys for reading hugetlb pool counters */
-enum { /* The number of pages of a given size that ... */
- HUGEPAGES_TOTAL, /* are allocated to the pool */
- HUGEPAGES_FREE, /* are not in use */
- HUGEPAGES_RSVD, /* are reserved for possible future use */
- HUGEPAGES_SURP, /* are allocated to the pool on demand */
- HUGEPAGES_OC, /* can be allocated on demand - maximum */
+enum { /* The number of pages of a given size that ... */
+ HUGEPAGES_TOTAL, /* are allocated to the pool */
+ HUGEPAGES_TOTAL_MEMPOL, /* are allocated following the NUMA mempolicy */
+ HUGEPAGES_FREE, /* are not in use */
+ HUGEPAGES_RSVD, /* are reserved for possible future use */
+ HUGEPAGES_SURP, /* are allocated to the pool on demand */
+ HUGEPAGES_OC, /* can be allocated on demand - maximum */
HUGEPAGES_MAX_COUNTERS,
};
#define get_huge_page_counter __pu_get_huge_page_counter
diff --git a/man/hugeadm.8 b/man/hugeadm.8
index 05cdceb..0c577bc 100644
--- a/man/hugeadm.8
+++ b/man/hugeadm.8
@@ -160,6 +160,14 @@ requested for the Minimum pool. The size of the pools
should be checked after
executing this command to ensure they were successful.
.TP
+.B --obey-numa-mempol
+
+This option requests that allocation of huge pages to the static pool with
+\fB--pool-pages-min\fP obey the NUMA memory policy of the current process. This
+policy can be explicitly specified using numactl or inherited from a parent
+process.
+
+.TP
.B --pool-pages-max=<size|DEFAULT>:[+|-]<pagecount|memsize<G|M|K>>
This option sets or adjusts the Maximum number of hugepages. Note that while
--
1.7.0.4
------------------------------------------------------------------------------
_______________________________________________
Libhugetlbfs-devel mailing list
Libhugetlbfs-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/libhugetlbfs-devel
