diff --git a/A3/execute_order.sh b/A3/execute_numa_test.sh
similarity index 76%
copy from A3/execute_order.sh
copy to A3/execute_numa_test.sh
index fde5d3b..265f5f2 100644
--- a/A3/execute_order.sh
+++ b/A3/execute_numa_test.sh
@@ -1,53 +1,56 @@
#!/bin/bash
#SBATCH --chdir /scratch/jaugey
#SBATCH --nodes 1
#SBATCH --ntasks 1
#SBATCH --cpus-per-task 28
#SBATCH --mem 10G
#SBATCH --partition serial
#SBATCH --account cs307
#SBATCH --reservation CS307-ex
-NAME=order
+NAME=numa_test
if [ -n "$1" ]; then
NAME=$1
fi
echo STARTING AT `date`
# local mode
echo "Local allocation"
numactl --localalloc ./$NAME
echo ""
# select node 0
echo "Binding node 0"
-numactl --membind=0 ./$NAME
+numactl --cpunodebind=1 --membind=0 ./$NAME
+numactl --cpunodebind=1 --membind=0 ./$NAME
+numactl --cpunodebind=1 --membind=0 ./$NAME
+numactl --cpunodebind=1 --membind=0 ./$NAME
echo ""
# select node 1
echo "Binding node 1"
numactl --membind=1 ./$NAME
echo ""
# interleaved mode
echo "Binding interleave node 0"
numactl --interleave=0 ./$NAME
echo ""
# interleaved mode
echo "Binding interleave node 1"
-numactl --interleave=1 ./$NAME
+numactl --interleave=0 ./$NAME
echo ""
# interleaved mode
echo "Binding interleave all nodes"
numactl --interleave=all ./$NAME
echo FINISHED at `date`
diff --git a/A3/execute_order.sh b/A3/execute_order.sh
index fde5d3b..83ae7e4 100644
--- a/A3/execute_order.sh
+++ b/A3/execute_order.sh
@@ -1,53 +1,30 @@
#!/bin/bash
#SBATCH --chdir /scratch/jaugey
#SBATCH --nodes 1
#SBATCH --ntasks 1
#SBATCH --cpus-per-task 28
#SBATCH --mem 10G
#SBATCH --partition serial
#SBATCH --account cs307
#SBATCH --reservation CS307-ex
NAME=order
if [ -n "$1" ]; then
NAME=$1
fi
echo STARTING AT `date`
# local mode
-echo "Local allocation"
-numactl --localalloc ./$NAME
+echo 1 Node
+numactl -C 0,1,2 ./$NAME
echo ""
-# select node 0
-echo "Binding node 0"
-numactl --membind=0 ./$NAME
+# separated node
+echo 2 Nodes
+numactl -C 0,1,27 ./$NAME
-echo ""
-
-# select node 1
-echo "Binding node 1"
-numactl --membind=1 ./$NAME
-
-echo ""
-
-# interleaved mode
-echo "Binding interleave node 0"
-numactl --interleave=0 ./$NAME
-
-echo ""
-
-# interleaved mode
-echo "Binding interleave node 1"
-numactl --interleave=1 ./$NAME
-
-echo ""
-
-# interleaved mode
-echo "Binding interleave all nodes"
-numactl --interleave=all ./$NAME
echo FINISHED at `date`
diff --git a/A3/execute_order.sh b/A3/execute_order_test.sh
similarity index 78%
copy from A3/execute_order.sh
copy to A3/execute_order_test.sh
index fde5d3b..7257cbf 100644
--- a/A3/execute_order.sh
+++ b/A3/execute_order_test.sh
@@ -1,53 +1,57 @@
#!/bin/bash
#SBATCH --chdir /scratch/jaugey
#SBATCH --nodes 1
#SBATCH --ntasks 1
#SBATCH --cpus-per-task 28
#SBATCH --mem 10G
#SBATCH --partition serial
#SBATCH --account cs307
#SBATCH --reservation CS307-ex
NAME=order
if [ -n "$1" ]; then
NAME=$1
fi
echo STARTING AT `date`
# local mode
echo "Local allocation"
numactl --localalloc ./$NAME
echo ""
# select node 0
echo "Binding node 0"
-numactl --membind=0 ./$NAME
+numactl --cpunodebind=1 --membind=0 ./$NAME
+numactl --cpunodebind=1 --membind=0 ./$NAME
+numactl --cpunodebind=1 --membind=0 ./$NAME
+numactl --cpunodebind=1 --membind=0 ./$NAME
+numactl --cpunodebind=1 --membind=0 ./$NAME
echo ""
# select node 1
echo "Binding node 1"
numactl --membind=1 ./$NAME
echo ""
# interleaved mode
echo "Binding interleave node 0"
numactl --interleave=0 ./$NAME
echo ""
# interleaved mode
echo "Binding interleave node 1"
numactl --interleave=1 ./$NAME
echo ""
# interleaved mode
echo "Binding interleave all nodes"
numactl --interleave=all ./$NAME
echo FINISHED at `date`
diff --git a/A3/numa.c b/A3/numa.c
index 8ec6d22..9772b70 100644
--- a/A3/numa.c
+++ b/A3/numa.c
@@ -1,135 +1,135 @@
#include <omp.h>
#include <stdio.h>
#include <stdlib.h>
#include "utility.h"
#include <stdint.h>
#define GB (uint64_t)(1 << 30)
#define SIZE (uint64_t)(8*GB)
volatile uint8_t *arr;
#define CACHE_LINE 64
// louis version accessor
/*inline uint64_t next_addr(uint64_t i){
// Change this part
if(i%3==0)
{
return 10 + 10*(arr[i]%10);
}
else
{
return 10 + 11*(arr[i]%11);
}
}*/
// louis version initializer
/*inline void init_array(rand_gen gen){
for(uint64_t i=0; i<SIZE; ++i){
arr[i] = (uint64_t)(next_rand(gen) * 10);
}
}*/
// ancarola version
inline uint64_t next_addr(uint64_t i){
return arr[i];
}
// 0, 1 alternated version
/*inline void init_array(rand_gen gen){
// elude linear pattern recognizer
for(uint64_t i=0; i<SIZE; i += 2 * CACHE_LINE){
uint64_t place = i / (2 * CACHE_LINE);
if (place % 2 == 0) {
// first place
arr[i] = 2 * CACHE_LINE + 1;
//printf("Allocando il dio boia: (%d, %d)\n", (int)i, (int)arr[i]);
} else {
// second place
arr[i+1] = 2 * CACHE_LINE - 1;
//printf("Allocando il dio boia: (%d, %d)\n", (int)(i+1), (int)arr[i+1]);
}
}
}*/
// linearly shifted version
/*inline void init_array(rand_gen gen){
// elude linear pattern recognizer
uint64_t counter = 0;
// shift 1 every two cache lines
for(uint64_t i=0; i<SIZE; i += 2 * CACHE_LINE){
uint64_t place = i / (2 * CACHE_LINE);
if (place % 2 == 0) {
arr[i + counter] = 2 * CACHE_LINE;
} else {
if ((counter + 1) % CACHE_LINE == 0) {
arr[i + counter] = CACHE_LINE + 1;
counter = 0;
} else {
arr[i + counter] = 2 * CACHE_LINE + 1;
++counter;
}
}
}
}*/
// random version
inline void init_array(rand_gen gen){
uint64_t r[2];
r[0] = 0;
r[1] = (uint64_t)(next_rand(gen) * (CACHE_LINE - 1));
- for(uint64_t i=0; i<SIZE; i += 4 * CACHE_LINE){
+ for(uint64_t i=0; i<SIZE; i += 2 * CACHE_LINE){
- arr[i + r[0]] = 4 * CACHE_LINE + (r[1] - r[0]);
+ arr[i + r[0]] = 2 * CACHE_LINE + (r[1] - r[0]);
r[0] = r[1];
r[1] = (uint64_t)(next_rand(gen) * (CACHE_LINE - 1));
}
}
int main(){
uint64_t i, counter;
double time;
volatile uint8_t temp;
arr = (uint8_t*) malloc(SIZE);
rand_gen gen = init_rand();
init_array(gen);
//Start timer
set_clock();
for(i=0, counter=0; i<SIZE; counter++){
temp = arr[i];
i += next_addr(i);
}
//Stop timer
time = elapsed_time();
temp = next_rand(gen)*10;
i = temp; // Just to suppress the compiler warning for not using temp and gen
if(counter < 10000000)
printf("ERROR: Too few accesses. You have to access more elements to measure reasonable time difference.\n");
if((time/counter) < 1.0e-8)
printf("ERROR: Time per access is too small. You have to further deoptimize the program to measure reasonable time difference.\n");
printf("Traversing %lx GB array took total time = %.4g seconds, number of accesses = %lu, %.4g seconds per access\n", SIZE/GB, time, counter, time/counter);
return 0;
}
diff --git a/A3/numa_test b/A3/numa_test
new file mode 100755
index 0000000..2641606
Binary files /dev/null and b/A3/numa_test differ
diff --git a/A3/numa.c b/A3/numa_test.c
similarity index 95%
copy from A3/numa.c
copy to A3/numa_test.c
index 8ec6d22..c5fc278 100644
--- a/A3/numa.c
+++ b/A3/numa_test.c
@@ -1,135 +1,150 @@
#include <omp.h>
#include <stdio.h>
#include <stdlib.h>
#include "utility.h"
#include <stdint.h>
#define GB (uint64_t)(1 << 30)
#define SIZE (uint64_t)(8*GB)
volatile uint8_t *arr;
#define CACHE_LINE 64
// louis version accessor
/*inline uint64_t next_addr(uint64_t i){
// Change this part
if(i%3==0)
{
return 10 + 10*(arr[i]%10);
}
else
{
return 10 + 11*(arr[i]%11);
}
}*/
// louis version initializer
/*inline void init_array(rand_gen gen){
for(uint64_t i=0; i<SIZE; ++i){
arr[i] = (uint64_t)(next_rand(gen) * 10);
}
}*/
// ancarola version
-inline uint64_t next_addr(uint64_t i){
+/*inline uint64_t next_addr(uint64_t i){
return arr[i];
+}*/
+
+inline uint64_t next_addr(uint64_t i){
+ return 1;
+}
+
+inline void init_array(rand_gen gen){
+ for(uint64_t i=0; i<SIZE; i++){
+ arr[i] = 1;
+ }
}
+
+
// 0, 1 alternated version
/*inline void init_array(rand_gen gen){
// elude linear pattern recognizer
for(uint64_t i=0; i<SIZE; i += 2 * CACHE_LINE){
uint64_t place = i / (2 * CACHE_LINE);
if (place % 2 == 0) {
// first place
arr[i] = 2 * CACHE_LINE + 1;
//printf("Allocando il dio boia: (%d, %d)\n", (int)i, (int)arr[i]);
} else {
// second place
arr[i+1] = 2 * CACHE_LINE - 1;
//printf("Allocando il dio boia: (%d, %d)\n", (int)(i+1), (int)arr[i+1]);
}
}
}*/
// linearly shifted version
/*inline void init_array(rand_gen gen){
// elude linear pattern recognizer
uint64_t counter = 0;
// shift 1 every two cache lines
for(uint64_t i=0; i<SIZE; i += 2 * CACHE_LINE){
uint64_t place = i / (2 * CACHE_LINE);
if (place % 2 == 0) {
arr[i + counter] = 2 * CACHE_LINE;
} else {
if ((counter + 1) % CACHE_LINE == 0) {
arr[i + counter] = CACHE_LINE + 1;
counter = 0;
} else {
arr[i + counter] = 2 * CACHE_LINE + 1;
++counter;
}
}
}
}*/
// random version
-inline void init_array(rand_gen gen){
+/*inline void init_array(rand_gen gen){
uint64_t r[2];
r[0] = 0;
r[1] = (uint64_t)(next_rand(gen) * (CACHE_LINE - 1));
for(uint64_t i=0; i<SIZE; i += 4 * CACHE_LINE){
arr[i + r[0]] = 4 * CACHE_LINE + (r[1] - r[0]);
r[0] = r[1];
r[1] = (uint64_t)(next_rand(gen) * (CACHE_LINE - 1));
}
-}
+}*/
+
+
+
int main(){
uint64_t i, counter;
double time;
volatile uint8_t temp;
arr = (uint8_t*) malloc(SIZE);
rand_gen gen = init_rand();
init_array(gen);
//Start timer
set_clock();
for(i=0, counter=0; i<SIZE; counter++){
temp = arr[i];
i += next_addr(i);
}
//Stop timer
time = elapsed_time();
temp = next_rand(gen)*10;
i = temp; // Just to suppress the compiler warning for not using temp and gen
if(counter < 10000000)
printf("ERROR: Too few accesses. You have to access more elements to measure reasonable time difference.\n");
if((time/counter) < 1.0e-8)
printf("ERROR: Time per access is too small. You have to further deoptimize the program to measure reasonable time difference.\n");
printf("Traversing %lx GB array took total time = %.4g seconds, number of accesses = %lu, %.4g seconds per access\n", SIZE/GB, time, counter, time/counter);
return 0;
}
diff --git a/A3/order_fence.c.save b/A3/order_fence.c.save
new file mode 100644
index 0000000..e69de29
diff --git a/A3/order_fence.c.save.1 b/A3/order_fence.c.save.1
new file mode 100644
index 0000000..fafbc01
--- /dev/null
+++ b/A3/order_fence.c.save.1
@@ -0,0 +1,80 @@
+#define _GNU_SOURCE
+#include <pthread.h>
+#include <omp.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdint.h>
+#include <stdbool.h>
+
+volatile int X, Y, r1, r2;
+volatile bool t1block, t2block, t1fin, t2fin;
+
+void *thread1Func(void *param){
+ for(;;){
+ while(t1block); // Wait for blocker to be removed
+ t1block = true; // Set up blocker
+
+ X = 1;
+ asm volatile("mfence" ::: "memory"); // Prevent any compiler reordering
+ r1 = Y;
+
+ t1fin = true; // Signal iteration end
+ }
+ return NULL; // Never returns
+}
+
+void *thread2Func(void *param){
+ for(;;){
+ while(t2block); // Wait for blocker to be removed
+ t2block = true; // Set up blocker
+
+ Y = 1;
+ asm volatile("mfence" ::: "memory"); // Prevent any compiler reordering
+ r2 = X;
+
+ t2fin = true; // Signal iteration end
+ }
+ return NULL; // Never returns
+}
+
+int main(){
+ //Init
+ t1block = t2block = true; // Act as blockers for threads
+ t1fin = t2fin = false; // Act as finish signals for threads
+
+ // Spawn the threads
+ pthread_t thread0, thread1, thread2;
+ thread0 = pthread_self();
+ pthread_create(&thread1, NULL, thread1Func, NULL);
+ pthread_create(&thread2, NULL, thread2Func, NULL);
+
+ /* Set the main thread affinity to CPU 0 */
+ cpu_set_t cpuset;
+ CPU_ZERO(&cpuset);
+ CPU_SET(0, &cpuset);
+ int s = pthread_setaffinity_np(thread0, sizeof(cpu_set_t), &cpuset);
+ if (s != 0){
+ printf("Failed to set main thread affinity\n");
+ return 0;
+ }
+
+ // Repeat the experiment
+ int detected = 0, i = 1;
+ for (i = 1; i < 1000000; i++){
+ // Reset all variables
+ X = 0;
+ Y = 0;
+ t1fin = t2fin = false;
+ t1block = t2block = false;
+
+ // Wait for threads to complete their iteration
+ while(!(t1fin && t2fin));
+
+ // Check if there was a memory reordering
+ if (r1 == 0 && r2 == 0){
+ detected++;
+ }
+ }
+ printf("%d memory re-orderings detected in %d iterations - %d percent\n", detected, i, 100*detected/i);
+ return 0; // Never returns
+}