mcexec: memory policy control by environmental variable

Refs: #1470
Change-Id: I3d556cae90d31d81572b1c4e5c680e826577d428

executer/include/defs.h

@@ -4671,7 +4671,7 @@ void cmd_ipcs(void);         /* ipcs.c */
 /*
  *  main.c
  */
-void main_loop(void);
+//void main_loop(void);
 void exec_command(void);
 struct command_table_entry *get_command_table_entry(char *);
 void program_usage(int);

executer/include/uprotocol.h

@@ -113,6 +113,18 @@ typedef unsigned long __cpu_set_unit;
 #define MPOL_NO_BSS               0x04
 #define MPOL_SHM_PREMAP           0x08
 
+/* should be the same as process.h */
+#define PLD_PROCESS_NUMA_MASK_BITS 256
+
+enum {
+	PLD_MPOL_DEFAULT,
+	PLD_MPOL_PREFERRED,
+	PLD_MPOL_BIND,
+	PLD_MPOL_INTERLEAVE,
+	PLD_MPOL_LOCAL,
+	PLD_MPOL_MAX,	/* always last member of enum */
+};
+
 #define PLD_MAGIC 0xcafecafe44332211UL
 
 struct program_load_desc {
@@ -147,6 +159,10 @@ struct program_load_desc {
 	unsigned long heap_extension;
 	long stack_premap;
 	unsigned long mpol_bind_mask;
+	int mpol_mode;
+	unsigned long mpol_nodemask[PLD_PROCESS_NUMA_MASK_BITS /
+				    (sizeof(unsigned long) * 8)];
+
 	int thp_disable;
 	int uti_thread_rank; /* N-th clone() spawns a thread on Linux CPU */
 	int uti_use_last_cpu; /* Work-around not to share CPU with OpenMP thread */

executer/user/mcexec.c

@@ -84,7 +84,11 @@
 #include "../include/pmi.h"
 #include "../include/qlmpi.h"
 #include <sys/xattr.h>
+#include "../include/defs.h"
 #include "../../lib/include/list.h"
+#include "../../lib/include/bitops-set_bit.h"
+#include "../../lib/include/bitops-clear_bit.h"
+#include "../../lib/include/bitops-test_bit.h"
 
 //#define DEBUG
 #define ADD_ENVS_OPTION
@@ -1056,6 +1060,64 @@ static inline cpu_set_t *numa_node_set(int n)
 	return (cpu_set_t *)(numa_nodes + n * cpu_set_size);
 }
 
+static inline void _numa_local(__cpu_set_unit *localset,
+			       unsigned long *nodemask, int nonlocal)
+{
+	int i;
+
+	memset(nodemask, 0, PLD_PROCESS_NUMA_MASK_BITS / 8);
+
+	for (i = 0; i < nnodes; i++) {
+		cpu_set_t *nodeset = numa_node_set(i);
+		int j;
+
+		if (nonlocal) {
+			set_bit(i, nodemask);
+		}
+
+		for (j = 0; j < ncpu; j++) {
+			if (test_bit(j, localset)) {
+				__dprintf("%d belongs to local set\n", j);
+			}
+
+			if (CPU_ISSET_S(j, cpu_set_size, nodeset)) {
+				__dprintf("%d belongs to node %d\n", j, i);
+			}
+
+			if (test_bit(j, localset) &&
+			    CPU_ISSET_S(j, cpu_set_size, nodeset)) {
+				if (nonlocal) {
+					clear_bit(i, nodemask);
+				} else {
+					set_bit(i, nodemask);
+				}
+			}
+		}
+	}
+}
+
+static inline void numa_local(__cpu_set_unit *localset, unsigned long *nodemask)
+{
+	_numa_local(localset, nodemask, 0);
+}
+
+static inline void numa_nonlocal(__cpu_set_unit *localset,
+				 unsigned long *nodemask)
+{
+	_numa_local(localset, nodemask, 1);
+}
+
+static inline void numa_all(unsigned long *nodemask)
+{
+	int i;
+
+	memset(nodemask, 0, PLD_PROCESS_NUMA_MASK_BITS / 8);
+
+	for (i = 0; i < nnodes; i++) {
+		set_bit(i, nodemask);
+	}
+}
+
 pid_t master_tid;
 
 pthread_mutex_t lock = PTHREAD_MUTEX_INITIALIZER;
@@ -2685,6 +2747,7 @@ int main(int argc, char **argv)
 	desc->heap_extension = heap_extension;
 
 	desc->mpol_bind_mask = 0;
+	desc->mpol_mode = PLD_MPOL_MAX; /* not specified */
 	if (mpol_bind_nodes) {
 		struct bitmask *bind_mask;
 		bind_mask = numa_parse_nodestring_all(mpol_bind_nodes);
@@ -2698,6 +2761,54 @@ int main(int argc, char **argv)
 			}
 		}
 	}
+	/* Fujitsu TCS specific: mempolicy */
+	else if (getenv("OMPI_MCA_plm_ple_memory_allocation_policy")) {
+		char *mpol =
+			getenv("OMPI_MCA_plm_ple_memory_allocation_policy");
+
+		__dprintf("OMPI_MCA_plm_ple_memory_allocation_policy: %s\n",
+			  mpol);
+
+		if (!strncmp(mpol, "localalloc", 10)) {
+			/* MPOL_DEFAULT has the same effect as MPOL_LOCAL */
+			desc->mpol_mode = MPOL_DEFAULT;
+		}
+		else if (!strncmp(mpol, "interleave_local", 16)) {
+			desc->mpol_mode = MPOL_INTERLEAVE;
+			numa_local(desc->cpu_set, desc->mpol_nodemask);
+		}
+		else if (!strncmp(mpol, "interleave_nonlocal", 19)) {
+			desc->mpol_mode = MPOL_INTERLEAVE;
+			numa_nonlocal(desc->cpu_set, desc->mpol_nodemask);
+		}
+		else if (!strncmp(mpol, "interleave_all", 14)) {
+			desc->mpol_mode = MPOL_INTERLEAVE;
+			numa_all(desc->mpol_nodemask);
+		}
+		else if (!strncmp(mpol, "bind_local", 10)) {
+			desc->mpol_mode = MPOL_BIND;
+			numa_local(desc->cpu_set, desc->mpol_nodemask);
+		}
+		else if (!strncmp(mpol, "bind_nonlocal", 13)) {
+			desc->mpol_mode = MPOL_BIND;
+			numa_nonlocal(desc->cpu_set, desc->mpol_nodemask);
+		}
+		else if (!strncmp(mpol, "bind_all", 8)) {
+			desc->mpol_mode = MPOL_BIND;
+			numa_all(desc->mpol_nodemask);
+		}
+		else if (!strncmp(mpol, "prefer_local", 12)) {
+			desc->mpol_mode = MPOL_PREFERRED;
+			numa_local(desc->cpu_set, desc->mpol_nodemask);
+		}
+		else if (!strncmp(mpol, "prefer_nonlocal", 15)) {
+			desc->mpol_mode = MPOL_PREFERRED;
+			numa_nonlocal(desc->cpu_set, desc->mpol_nodemask);
+		}
+
+		__dprintf("mpol_mode: %d, mpol_nodemask: %ld\n",
+			  desc->mpol_mode, desc->mpol_nodemask[0]);
+	}
 
 	desc->uti_thread_rank = uti_thread_rank;
 	desc->uti_use_last_cpu = uti_use_last_cpu;