Skip to content
GitLab
Projects
Groups
Snippets
/
Help
Help
Support
Community forum
Keyboard shortcuts
?
Submit feedback
Contribute to GitLab
Sign in
Toggle navigation
Menu
Open sidebar
Priyal Suneja
single-server-etrace
Commits
12dd7583
Commit
12dd7583
authored
Mar 06, 2022
by
Priyal Suneja
Browse files
created msr.c for signal stuff
parent
11225974
Changes
27
Expand all
Hide whitespace changes
Inline
Side-by-side
benchmarks/tests/msr/Makefile
View file @
12dd7583
...
...
@@ -3,7 +3,9 @@ BUILDDIR = build
#all: tlb_msr l2_msr l1_msr ins_msr ins l2 tlb l1 l1_populate l2_populate \
# tlb_populate
all
:
tlb_msr l1_msr l2_msr qs_msr
#all: tlb_msr l1_msr l2_msr qs_msr mm_msr mm_signal
all
:
mm_signal
l1_msr
:
l1_msr.c utils.c msr.c
gcc
-O0
-Wall
-o
$(BUILDDIR)
/l1_msr l1_msr.c utils.c msr.c
-lm
...
...
@@ -32,6 +34,9 @@ qs_msr: qs_msr.c msr.c utils.c
mm_msr
:
mm_msr.c msr.c utils.c
gcc
-O0
-g
-Wall
-o
$(BUILDDIR)
/mm_msr mm_msr.c utils.c msr.c
-lm
mm_signal
:
mm_signal.c msr.c utils.c
gcc
-O0
-g
-Wall
-o
$(BUILDDIR)
/mm_signal mm_signal.c utils.c msr.c
-lm
ins
:
ins.c msr.c utils.c
gcc
-O0
-Wall
-o
$(BUILDDIR)
/ins ins.c msr.c utils.c
-lm
...
...
benchmarks/tests/msr/build/mm_msr
0 → 100755
View file @
12dd7583
File added
benchmarks/tests/msr/build/mm_signal
0 → 100755
View file @
12dd7583
File added
benchmarks/tests/msr/mm_msr.c
View file @
12dd7583
...
...
@@ -32,14 +32,7 @@ int measure_msr(int cpu_model, int cpu_info[3], double energy_units[2],
long
long
result
;
double
package_before
,
package_after
;
double
dram_before
,
dram_after
;
int
one
[
MM_SIZE
][
MM_SIZE
],
two
[
MM_SIZE
][
MM_SIZE
],
res
[
MM_SIZE
][
MM_SIZE
];
for
(
int
i
=
0
;
i
<
MM_SIZE
;
i
++
)
{
for
(
int
j
=
0
;
j
<
MM_SIZE
;
j
++
)
{
one
[
i
][
j
]
=
i
*
j
;
two
[
i
][
j
]
=
i
+
j
;
}
}
// int one[MM_SIZE][MM_SIZE], two[MM_SIZE][MM_SIZE], res[MM_SIZE][MM_SIZE];
fd
=
open_msr
(
0
);
// todo: add package detection + map and stuff
...
...
@@ -58,7 +51,23 @@ int measure_msr(int cpu_model, int cpu_info[3], double energy_units[2],
close
(
fd
);
for
(
int
i
=
0
;
i
<
ITERATIONS_PER_RUN
;
i
++
)
{
int
**
one
=
malloc
(
MM_SIZE
*
sizeof
(
int
*
));
int
**
two
=
malloc
(
MM_SIZE
*
sizeof
(
int
*
));
int
**
res
=
malloc
(
MM_SIZE
*
sizeof
(
int
*
));
for
(
int
i
=
0
;
i
<
MM_SIZE
;
i
++
)
{
one
[
i
]
=
malloc
(
MM_SIZE
*
sizeof
(
int
));
two
[
i
]
=
malloc
(
MM_SIZE
*
sizeof
(
int
));
res
[
i
]
=
malloc
(
MM_SIZE
*
sizeof
(
int
));
}
for
(
int
i
=
0
;
i
<
MM_SIZE
;
i
++
)
{
for
(
int
j
=
0
;
j
<
MM_SIZE
;
j
++
)
{
one
[
i
][
j
]
=
i
*
j
;
two
[
i
][
j
]
=
i
+
j
;
}
}
for
(
int
i
=
0
;
i
<
ITERATIONS_PER_RUN2
;
i
++
)
{
matrix_multiply
(
one
,
two
,
res
);
}
...
...
benchmarks/tests/msr/mm_signal.c
0 → 100644
View file @
12dd7583
/*
* author: Priyal Suneja ; suneja@cs.washington.edu
*
* to run: sudo ./build/mm_msr
*/
#include
"msr.h"
// int cpu_info[3]; // 0 -> pp0, 1-> pp1, 2-> dram
// double energy_units[2]; // 0 -> cpu, 1 -> dram
// extern int errno;
// double readings[1000000];
// double reading;
// FILE *fptr;
// unsigned int count = 0;
// void sig_handler(int signum) {
// int fd=open_msr(0); // todo: add package detection + map and stuff
//
// /* Package Energy */
// long long result = read_msr(fd,MSR_PKG_ENERGY_STATUS);
//
// close(fd);
//
// reading = (double)result*energy_units[0];
// fprintf(fptr, "%f\n", reading);
//
// }
void
matrix_multiply
(
int
**
one
,
int
**
two
,
int
**
result
)
{
int
rows1
=
MM_SIZE
;
int
cols2
=
MM_SIZE
;
int
common
=
MM_SIZE
;
//cols of 1 and rows of 2
for
(
int
rr
=
0
;
rr
<
rows1
;
rr
++
)
{
for
(
int
rc
=
0
;
rc
<
cols2
;
rc
++
)
{
result
[
rr
][
rc
]
=
0
;
}
}
for
(
int
i
=
0
;
i
<
rows1
;
i
++
)
{
for
(
int
j
=
0
;
j
<
cols2
;
j
++
)
{
for
(
int
k
=
0
;
k
<
common
;
k
++
)
{
result
[
i
][
j
]
+=
one
[
i
][
k
]
*
two
[
k
][
j
];
}
}
}
}
void
matrix_multiply_measure
()
{
int
**
one
=
malloc
(
MM_SIZE
*
sizeof
(
int
*
));
int
**
two
=
malloc
(
MM_SIZE
*
sizeof
(
int
*
));
int
**
res
=
malloc
(
MM_SIZE
*
sizeof
(
int
*
));
for
(
int
i
=
0
;
i
<
MM_SIZE
;
i
++
)
{
one
[
i
]
=
malloc
(
MM_SIZE
*
sizeof
(
int
));
two
[
i
]
=
malloc
(
MM_SIZE
*
sizeof
(
int
));
res
[
i
]
=
malloc
(
MM_SIZE
*
sizeof
(
int
));
}
for
(
int
i
=
0
;
i
<
MM_SIZE
;
i
++
)
{
for
(
int
j
=
0
;
j
<
MM_SIZE
;
j
++
)
{
one
[
i
][
j
]
=
i
*
j
;
two
[
i
][
j
]
=
i
+
j
;
}
}
for
(
int
i
=
0
;
i
<
100
;
i
++
)
{
matrix_multiply
(
one
,
two
,
res
);
}
for
(
int
i
=
0
;
i
<
MM_SIZE
;
i
++
)
{
free
(
one
[
i
]);
free
(
two
[
i
]);
free
(
res
[
i
]);
}
free
(
one
);
free
(
two
);
free
(
res
);
return
;
}
int
main
(
int
argc
,
char
*
argv
[])
{
// fptr = fopen("mm_signal_out", "w+");
//
// if(!fptr) {
// printf("errno: %d ", errno);
// printf("err string: %s\n", strerror(errno));
// return 0;
// }
//
// signal(SIGALRM, sig_handler);
//
// ualarm(500*1000,500*1000); // 1000 us = 1 ms; 0.5s = 500 ms
// // ualarm(500,500);
//
// get_cpu_info(CPU_HASWELL_EP, cpu_info, energy_units);
//
// measure_msr();
//
// // fclose(fptr);
printf
(
"%s
\n
"
,
strcat
(
argv
[
0
],
"_out"
));
measure_msr
(
"mm_signal_out"
,
&
matrix_multiply_measure
);
return
0
;
}
benchmarks/tests/msr/mm_signal_out
0 → 100644
View file @
12dd7583
benchmarks/tests/msr/msr.c
View file @
12dd7583
...
...
@@ -3,6 +3,8 @@
*/
#include
"msr.h"
extern
int
errno
;
int
open_msr
(
int
core
)
{
char
msr_filename
[
BUFSIZ
];
...
...
@@ -42,7 +44,6 @@
int
get_cpu_info
(
int
cpu_model
,
int
cpu_info
[
3
],
double
energy_units
[
2
])
{
int
fd
,
result
;
// double power_units,time_units;
double
cpu_energy_units
,
dram_energy_units
;
/** indices to cpu_info **/
...
...
@@ -110,32 +111,48 @@ int get_cpu_info(int cpu_model, int cpu_info[3], double energy_units[2]) {
}
// printf("\tListing paramaters for package #0\n");
fd
=
open_msr
(
0
);
// todo: add package detection + map and stuff
/* Calculate the units used */
result
=
read_msr
(
fd
,
MSR_RAPL_POWER_UNIT
);
// power_units=pow(0.5,(double)(result&0xf));
cpu_energy_units
=
pow
(
0
.
5
,(
double
)((
result
>>
8
)
&
0x1f
));
// time_units=pow(0.5,(double)((result>>16)&0xf));
/* On Haswell EP and Knights Landing */
/* The DRAM units differ from the CPU ones */
dram_energy_units
=
pow
(
0
.
5
,(
double
)
16
);
// we know we're on haswell
// so this is ok
// printf("DRAM: Using %lf instead of %lf\n",
// dram_energy_units,cpu_energy_units);
//
// printf("\t\tPower units = %.3fW\n",power_units);
// printf("\t\tCPU Energy units = %.8fJ\n",cpu_energy_units);
// printf("\t\tDRAM Energy units = %.8fJ\n",dram_energy_units);
// printf("\t\tTime units = %.8fs\n",time_units);
energy_units
[
0
]
=
cpu_energy_units
;
energy_units
[
1
]
=
dram_energy_units
;
// printf("\n");
return
0
;
}
void
sig_handler
(
int
signum
)
{
int
fd
=
open_msr
(
0
);
long
long
result
=
read_msr
(
fd
,
MSR_PKG_ENERGY_STATUS
);
close
(
fd
);
reading
=
(
double
)
result
*
energy_units
[
0
];
fprintf
(
fptr
,
"%f
\n
"
,
reading
);
}
void
measure_msr
(
char
*
filename
,
void
(
*
func_ptr
)())
{
fptr
=
fopen
(
filename
,
"w+"
);
if
(
!
fptr
)
{
printf
(
"errno: %d, err_string: %s
\n
"
,
errno
,
strerror
(
errno
));
return
;
}
signal
(
SIGALRM
,
sig_handler
);
ualarm
(
500
*
1000
,
500
*
1000
);
// 1000 us = 1ms,; 0.5s = 500 ms
get_cpu_info
(
CPU_HASWELL_EP
,
cpu_info
,
energy_units
);
func_ptr
();
}
benchmarks/tests/msr/msr.h
View file @
12dd7583
...
...
@@ -2,6 +2,7 @@
* author: Priyal Suneja ; suneja@cs.washington.edu
*/
#include
<stdio.h>
#include
<signal.h>
#include
<string.h>
#include
<stdlib.h>
#include
<unistd.h>
...
...
@@ -15,7 +16,8 @@
#include
"utils.h"
#define RUNS 1
#define ITERATIONS_PER_RUN 1000
// #define ITERATIONS_PER_RUN 1000
#define ITERATIONS_PER_RUN 100
#define L1_SIZE 32*1024
#define L2_SIZE 256*1024
#define PAGE_SIZE 4*1024
...
...
@@ -24,7 +26,8 @@
#define L1_LL_SIZE 5*5*L1_SIZE
#define L2_LL_SIZE 5*5*L2_SIZE
#define TLB_LL_SIZE TLB_ASSOC*TLB_ENTRIES*8*10
#define MM_SIZE L1_SIZE/4
// #define MM_SIZE L1_SIZE/4
#define MM_SIZE 1024
#define MSR_RAPL_POWER_UNIT 0x606
...
...
@@ -98,4 +101,11 @@
int
open_msr
(
int
core
);
long
long
read_msr
(
int
fd
,
int
which
);
void
sig_handler
(
int
signum
);
int
get_cpu_info
(
int
cpu_model
,
int
cpu_info
[
3
],
double
energy_units
[
2
]);
void
measure_msr
(
char
*
filename
,
void
(
*
funcptr
)());
int
cpu_info
[
3
];
double
energy_units
[
2
];
double
reading
;
FILE
*
fptr
;
benchmarks/tests/msr/out
0 → 100644
View file @
12dd7583
This diff is collapsed.
Click to expand it.
data/bc
View file @
12dd7583
********* bc *********
energy consumed: 16.
270386
energy consumed: 16.
436829
Performance counter stats for '/homes/sys/suneja/treehouse/single-server-etrace/benchmarks/gapbs/build/bc -g 10 -n 1':
1,
317,513
L1-icache-load-misses (
39.97
%)
5,
748,132,504
cycles (5
3
.0
0
%)
3,7
80,595,818
cycle_activity.cycles_no_execute (54.4
0
%)
2,26
5
,9
8
4,
682
instructions # 0.
39
insn per cycle (6
7.12
%)
1,
140,725
l2_rqsts.miss (6
8.55
%)
336,691
dTLB-load-misses (
70.03
%)
2,3
43
,90
1
L1-dcache-load-misses (70.
2
0%)
33,524
iTLB-load-misses (55.
95
%)
2,264
LLC-load-misses (4
0.90
%)
1,
699,080
L1-icache-load-misses (
40.78
%)
5,
699,354,075
cycles (5
4
.0
8
%)
3,7
61,727,166
cycle_activity.cycles_no_execute (54.
6
4%)
2,26
0
,9
9
4,
450
instructions # 0.
40
insn per cycle (6
6.34
%)
1,
212,829
l2_rqsts.miss (6
7.88
%)
200,202
dTLB-load-misses (
69.65
%)
2,3
68
,90
3
L1-dcache-load-misses (70.
1
0%)
27,156
iTLB-load-misses (55.
16
%)
1,860
LLC-load-misses (4
1.78
%)
0.16
4407234
seconds time elapsed
0.16
6086945
seconds time elapsed
data/bfs
View file @
12dd7583
********* bfs *********
energy consumed: 1
3.098206
energy consumed: 1
2.690674
Performance counter stats for '/homes/sys/suneja/treehouse/single-server-etrace/benchmarks/gapbs/build/bfs -g 10 -n 1':
986,953
L1-icache-load-misses (
40.44
%)
4,
462,342,907
cycles (51.
92
%)
2,
945,300,555
cycle_activity.cycles_no_execute (5
4.24
%)
1,7
65,950,369
instructions # 0.40 insn per cycle (6
5.56
%)
9
86,530
l2_rqsts.miss (67.
86
%)
241,789
dTLB-load-misses (70.
7
1%)
1,
680,535
L1-dcache-load-misses (7
0.15
%)
25,836
iTLB-load-misses (5
5.53
%)
728
LLC-load-misses (41.
07
%)
1,107,304
L1-icache-load-misses (
37.56
%)
4,
348,369,194
cycles (51.
31
%)
2,
854,498,298
cycle_activity.cycles_no_execute (5
3.85
%)
1,7
24,011,274
instructions # 0.40 insn per cycle (6
6.89
%)
9
61,015
l2_rqsts.miss (67.
71
%)
157,857
dTLB-load-misses (70.
3
1%)
1,
537,067
L1-dcache-load-misses (7
2.64
%)
19,757
iTLB-load-misses (5
6.62
%)
1,692
LLC-load-misses (41.
31
%)
0.13
5875754
seconds time elapsed
0.13
0066105
seconds time elapsed
data/cc
View file @
12dd7583
********* cc *********
energy consumed: 1
0.851562
energy consumed: 1
1.344421
Performance counter stats for '/homes/sys/suneja/treehouse/single-server-etrace/benchmarks/gapbs/build/cc -g 10 -n 1':
521,
83
0
L1-icache-load-misses (3
3.96
%)
3,94
3,950,262
cycles (
46.02
%)
2,
571,714,158
cycle_activity.cycles_no_execute (5
0.22
%)
1,5
48,869,904
instructions # 0.
39
insn per cycle (6
4.54
%)
703,233
l2_rqsts.miss (
69.13
%)
1
59,765
dTLB-load-misses (7
3.63
%)
1,
146,086
L1-dcache-load-misses (7
5.84
%)
1
3,230
iTLB-load-misses (5
8.71
%)
2,052
LLC-load-misses (3
8.51
%)
1,190,5
83 L1-icache-load-misses (3
8.62
%)
3,94
6,377,568
cycles (
51.91
%)
2,
625,438,509
cycle_activity.cycles_no_execute (5
4.85
%)
1,5
70,912,417
instructions # 0.
40
insn per cycle (6
7.29
%)
896,155
l2_rqsts.miss (
71.48
%)
1
48,057
dTLB-load-misses (7
2.99
%)
1,
657,101
L1-dcache-load-misses (7
0.98
%)
1
9,642
iTLB-load-misses (5
4.30
%)
231
LLC-load-misses (3
6.77
%)
0.11
1131435
seconds time elapsed
0.11
6891287
seconds time elapsed
data/cc_sv
View file @
12dd7583
********* cc_sv *********
energy consumed: 12.
536682
energy consumed: 12.
314270
Performance counter stats for '/homes/sys/suneja/treehouse/single-server-etrace/benchmarks/gapbs/build/cc_sv -g 10 -n 1':
822,402
L1-icache-load-misses (4
0.98
%)
4,29
7,061,655
cycles (53.
90
%)
2,8
44,930,438
cycle_activity.cycles_no_execute (55.
68
%)
1,
712,174,504
instructions # 0.40 insn per cycle (6
8.32
%)
872,453
l2_rqsts.miss (6
9.21
%)
168,385
dTLB-load-misses (
70.41
%)
1,
247,252
L1-dcache-load-misses (
69.08
%)
17,790
iTLB-load-misses (5
4.15
%)
1,500
LLC-load-misses (4
0.48
%)
1,033,315
L1-icache-load-misses (4
1.40
%)
4,29
0,480,362
cycles (53.
54
%)
2,8
15,212,359
cycle_activity.cycles_no_execute (55.
04
%)
1,
697,645,725
instructions # 0.40 insn per cycle (6
5.90
%)
962,305
l2_rqsts.miss (6
6.54
%)
220,918
dTLB-load-misses (
67.76
%)
1,
669,248
L1-dcache-load-misses (
70.57
%)
25,839
iTLB-load-misses (5
5.69
%)
449
LLC-load-misses (4
3.00
%)
0.12
8010521
seconds time elapsed
0.12
6721249
seconds time elapsed
data/err_bfs_10
View file @
12dd7583
********* graph500 *********
energy consumed:
11592.145508
energy consumed:
7315.767639
Performance counter stats for 'mpirun --allow-run-as-root --mca orte_base_help_aggregate 0 /homes/sys/suneja/treehouse/single-server-etrace/benchmarks/graph500/build/err_bfs 10':
9
8,
699,460
L1-icache-load-misses (44.44%)
2,332,035,353,759
cycles (55.5
5
%)
1,152,715,011,364
cycle_activity.cycles_no_execute (55.5
5
%)
1,
850,752,817,861
instructions # 0.
79
insn per cycle (66.6
7
%)
98,191,039
l2_rqsts.miss (66.6
7
%)
7,238,790
dTLB-load-misses (66.67%)
258,107,08
1 L1-dcache-load-misses (66.6
7
%)
4,242,931
iTLB-load-misses (55.5
5
%)
49
1,0
02
LLC-load-misses (44.4
4
%)
6
8,
218,274
L1-icache-load-misses (44.44%)
1,474,840,161,372
cycles (55.5
4
%)
728,056,171,420
cycle_activity.cycles_no_execute (55.5
3
%)
1,
181,545,305,536
instructions # 0.
80
insn per cycle (66.6
5
%)
74,692,596
l2_rqsts.miss (66.6
6
%)
5,435,322
dTLB-load-misses (66.67%)
162,300,96
1 L1-dcache-load-misses (66.6
8
%)
2,923,076
iTLB-load-misses (55.5
7
%)
66
1,0
43
LLC-load-misses (44.4
5
%)
111.573353900
seconds time elapsed
70.666371126
seconds time elapsed
data/err_bfs_5
View file @
12dd7583
********* graph500 *********
energy consumed:
6648.644470
energy consumed:
-261653.963745
Performance counter stats for 'mpirun --allow-run-as-root --mca orte_base_help_aggregate 0 /homes/sys/suneja/treehouse/single-server-etrace/benchmarks/graph500/build/err_bfs 5':
57,714,008
L1-icache-load-misses (44.
44
%)
1,339,595,077,610
cycles (55.5
4
%)
665,801,528,036
cycle_activity.cycles_no_execute (55.
5
4%)
1,051,534,441,842
instructions # 0.7
8
insn per cycle (66.
66
%)
52,637,068
l2_rqsts.miss (66.
67
%)
4,428,000
dTLB-load-misses (66.
67
%)
131,650
,8
4
3 L1-dcache-load-misses (66.
6
7%)
2,631,684
iTLB-load-misses (55.
5
7%)
4
91,
02
4
LLC-load-misses (44.
44
%)
14,826,075
L1-icache-load-misses (44.
31
%)
105,809,742,622
cycles (55.5
1
%)
52,828,424,983
cycle_activity.cycles_no_execute (55.
6
4%)
81,347,638,321
instructions # 0.7
7
insn per cycle (66.
82
%)
33,051,302
l2_rqsts.miss (66.
88
%)
2,200,341
dTLB-load-misses (66.
89
%)
44,803
,8
1
3 L1-dcache-load-misses (66.7
2
%)
995,298
iTLB-load-misses (55.
3
7%)
4
77,4
02 LLC-load-misses (44.
19
%)
64.285067760
seconds time elapsed
5.386520128
seconds time elapsed
data/err_bfs_7
View file @
12dd7583
********* graph500 *********
energy consumed:
6912.886353
energy consumed:
2162.643616
Performance counter stats for 'mpirun --allow-run-as-root --mca orte_base_help_aggregate 0 /homes/sys/suneja/treehouse/single-server-etrace/benchmarks/graph500/build/err_bfs 7':
61,432,088
L1-icache-load-misses (44.44%)
1,388,704,608,402
cycles (55.5
4
%)
688,698,870,707
cycle_activity.cycles_no_execute (55.5
3
%)
1,099,346,02
0,0
09
instructions # 0.79 insn per cycle (66.6
5
%)
5
7,
035,867
l2_rqsts.miss (66.6
6
%)
3,
720,605
dTLB-load-misses (66.6
7
%)
166,557,847
L1-dcache-load-misses (66.6
8
%)
3
,5
9
7,
340
iTLB-load-misses (55.5
8
%)
660,096
LLC-load-misses (44.4
5
%)
29,166,265
L1-icache-load-misses (44.44%)
439,278,004,248
cycles (55.5
5
%)
216,548,398,356
cycle_activity.cycles_no_execute (55.5
4
%)
348,518,35
0,0
83
instructions # 0.79 insn per cycle (66.6
7
%)
3
7,
689,328
l2_rqsts.miss (66.6
7
%)
3,
144,280
dTLB-load-misses (66.6
8
%)
86,998,922
L1-dcache-load-misses (66.6
7
%)
1
,5
3
7,
667
iTLB-load-misses (55.5
7
%)
351,358
LLC-load-misses (44.4
3
%)
66.593541155
seconds time elapsed
21.324661887
seconds time elapsed
data/err_sssp_10
View file @
12dd7583
********* graph500 *********
energy consumed:
55534.383545
energy consumed:
30378.670776
Performance counter stats for 'mpirun --allow-run-as-root --mca orte_base_help_aggregate 0 /homes/sys/suneja/treehouse/single-server-etrace/benchmarks/graph500/build/err_sssp 10':
433,632,630
L1-icache-load-misses (44.44%)
11,135,824,616,281
cycles (55.5
6
%)
5,473,242,088,074
cycle_activity.cycles_no_execute (55.5
6
%)
9,000,792,212,049
instructions # 0.81 insn per cycle (66.67%)
2
75,
758,467
l2_rqsts.miss (66.6
7
%)
1
9,917,309
dTLB-load-misses (66.6
7
%)
979,053,106
L1-dcache-load-misses (66.67%)
1
8,
082,9
53 iTLB-load-misses (55.5
5
%)
505,250
LLC-load-misses (44.4
4
%)
247,144,091
L1-icache-load-misses (44.44%)
6,093,495,981,209
cycles (55.5
5
%)
2,995,936,592,528
cycle_activity.cycles_no_execute (55.5
5
%)
4,934,513,027,092
instructions # 0.81 insn per cycle (66.67%)
1
75,
090,009
l2_rqsts.miss (66.6
6
%)
1
4,607,687
dTLB-load-misses (66.6
6
%)
453,875,342
L1-dcache-load-misses (66.67%)
8,
290,0
53 iTLB-load-misses (55.5
6
%)
693,911
LLC-load-misses (44.4
5
%)
531.1891313
56 seconds time elapsed
290.9186372
56 seconds time elapsed
data/err_sssp_5
View file @
12dd7583
********* graph500 *********
energy consumed:
1392.513306
energy consumed:
9996.351929
Performance counter stats for 'mpirun --allow-run-as-root --mca orte_base_help_aggregate 0 /homes/sys/suneja/treehouse/single-server-etrace/benchmarks/graph500/build/err_sssp 5':
24,555,265
L1-icache-load-misses (44.4
5
%)
285,392,534,393
cycles (55.55%)
142,691,242,015
cycle_activity.cycles_no_execute (55.5
1
%)
216,933,940,444
instructions # 0.7
6
insn per cycle (66.6
4
%)
35,934,839
l2_rqsts.miss (66.6
3
%)
2,241,228
dTLB-load-misses (66.6
4
%)
65,650,795
L1-dcache-load-misses (66.6
8
%)
1
,3
35,792
iTLB-load-misses (55.6
2
%)
526,891
LLC-load-misses (44.4
8
%)
85,304,279
L1-icache-load-misses (44.4
4
%)
2,007,011,249,671
cycles (55.55%)
995,820,691,821
cycle_activity.cycles_no_execute (55.5
6
%)
1,578,038,456,142
instructions # 0.7
9
insn per cycle (66.6
7
%)
69,862,772
l2_rqsts.miss (66.6
7
%)
4,793,871
dTLB-load-misses (66.6
7
%)
202,389,472
L1-dcache-load-misses (66.6
7
%)
3
,3
96,159
iTLB-load-misses (55.
5
6%)
606,285
LLC-load-misses (44.4
4
%)
13.93833649
4 seconds time elapsed
96.04620810
4 seconds time elapsed
data/err_sssp_7
View file @
12dd7583
********* graph500 *********
energy consumed:
25180.204590
energy consumed:
39621.261475
Performance counter stats for 'mpirun --allow-run-as-root --mca orte_base_help_aggregate 0 /homes/sys/suneja/treehouse/single-server-etrace/benchmarks/graph500/build/err_sssp 7':
198,047,737
L1-icache-load-misses (44.4
4
%)
5,052,489,829,550
cycles (55.5
5
%)
2,490,755,529,486
cycle_activity.cycles_no_execute (55.5
6
%)
4,037,942,862,600
instructions # 0.8
0
insn per cycle (66.67%)
1
30,438,276
l2_rqsts.miss (66.67%)
8,730,454
dTLB-load-misses (66.6
7
%)
390,230
,28
6
L1-dcache-load-misses (66.67%)
6,511,198
iTLB-load-misses (55.56%)
3
84,
875
LLC-load-misses (44.4
4
%)
305,617,322
L1-icache-load-misses (44.4
5
%)
7,941,593,121,825
cycles (55.5
6
%)
3,913,962,190,944
cycle_activity.cycles_no_execute (55.5
5
%)
6,396,995,410,148
instructions # 0.8
1
insn per cycle (66.67%)
1
83,576,848
l2_rqsts.miss (66.67%)
12,006,762
dTLB-load-misses (66.6
6
%)
887,132
,28
4
L1-dcache-load-misses (66.67%)
12,268,891
iTLB-load-misses (55.56%)
5
84,
500
LLC-load-misses (44.4
5
%)
241.252330907
seconds time elapsed
378.973453342
seconds time elapsed
data/l1_msr
View file @
12dd7583
*******************************Average over 1 runs: 2
08.768921
*******************************
*******************************Average over 1 runs: 2
30.815369
*******************************
Performance counter stats for '/homes/sys/suneja/treehouse/single-server-etrace/benchmarks/tests/msr/build/l1_msr':
1
78,414
L1-icache-load-misses (44.4
5
%)
6
,5
46,702,787
cycles (55.
56
%)
1,
376,557,774
cycle_activity.cycles_no_execute (55.
56
%)
6,00
6
,3
79,928
instructions # 0.9
2
insn per cycle (66.
67
%)
69,852,68
2 l2_rqsts.miss (66.
6
7%)
123,396
dTLB-load-misses (66.
67
%)
81
7,018,43
5 L1-dcache-load-misses (66.
66
%)
13,969
iTLB-load-misses (55.
55
%)
7
LLC-load-misses (44.
44
%)
1
25,590
L1-icache-load-misses (44.
1
4%)
7
,5
59,444,092
cycles (55.
31
%)
1,
808,223,419
cycle_activity.cycles_no_execute (55.
42
%)
6,00
9
,3
41,465
instructions # 0.
7
9 insn per cycle (66.
59
%)
56,604,71
2 l2_rqsts.miss (66.7
4
%)
3,953,061
dTLB-load-misses (66.
89
%)
81
9,630,80
5 L1-dcache-load-misses (66.
93
%)
270
iTLB-load-misses (55.
60
%)
2,572
LLC-load-misses (44.
28
%)