I recently needed a way to measure how much energy a piece of code consumes. This is easy to do in Linux using the perf utility assuming the underlying kernel has been built with appropriate flags. This looks something like this
$ sudo perf stat -a -e power/energy-pkg/ ls -la /usr
total 116
drwxr-xr-x 14 root root 4096 Mar 28 02:08 .
drwxr-xr-x 19 root root 4096 Apr 16 16:02 ..
drwxr-xr-x 2 root root 36864 May 7 15:05 bin
drwxr-xr-x 2 root root 4096 Apr 18 2022 games
drwxr-xr-x 37 root root 16384 May 7 15:51 include
drwxr-xr-x 88 root root 4096 May 7 15:05 lib
drwxr-xr-x 2 root root 4096 Mar 28 02:08 lib32
drwxr-xr-x 2 root root 4096 Apr 19 06:14 lib64
drwxr-xr-x 9 root root 4096 Apr 16 18:02 libexec
drwxr-xr-x 2 root root 4096 Mar 28 02:08 libx32
drwxr-xr-x 10 root root 4096 Mar 28 02:08 local
drwxr-xr-x 2 root root 20480 May 7 15:20 sbin
drwxr-xr-x 119 root root 4096 May 7 15:05 share
drwxr-xr-x 7 root root 4096 Apr 20 06:56 src
Performance counter stats for 'system wide':
0.03 Joules power/energy-pkg/
0.000777146 seconds time elapsed
This tells me that the ls command consumed 0.03 Joules of energy and ran for 0.00078 seconds. I needed to get similar data for a C function that might be surrounded by code that I do not need to measure. Something like this:
start_measurement();
function_call_one();
end_measurement();
start_measurement();
function_call_two();
end_measurement();
Here the two function calls need to be in the same translation unit, so I cannot separate them and then use the perf tool to measure energy consumption.
I started with looking at the source code for perf-stat but I was lost! This command takes in a whole bunch of options, the code was hard to follow. I googled How perf works and found Julia Evans’ blog post titled exactly that. That post pointed me to the system call perf_event_open. The man page has an example code at the end which gets total number of instructions executed for a printf call, while that is useful I still do not know what particular parameters I need to use to get energy info. At this point I decided to strace the perf stat command to see how it was using perf_event_open.
$ sudo strace -v -s 100000 perf stat -a -e power/energy-pkg/ ls -la /usr > output.txt 2>&1
Grepping for the word perf in output.txt revealed three calls to perf_event_open
$ rg perf_event_open output.txt
2161:perf_event_open({type=0x17 /* PERF_TYPE_??? */, size=PERF_ATTR_SIZE_VER7, config=0x2, sample_period=0, sample_type=PERF_SAMPLE_IDENTIFIER, read_format=PERF_FORMAT_TOTAL_TIME_ENABLED|PERF_FORMAT_TOTAL_TIME_RUNNING, disabled=1, inherit=1, pinned=0, exclusive=0, exclude_user=0, exclude_kernel=0, exclude_hv=0, exclude_idle=0, mmap=0, comm=0, freq=0, inherit_stat=0, enable_on_exec=0, task=0, watermark=0, precise_ip=0 /* arbitrary skid */, mmap_data=0, sample_id_all=0, exclude_host=0, exclude_guest=1, exclude_callchain_kernel=0, exclude_callchain_user=0, mmap2=0, comm_exec=0, use_clockid=0, context_switch=0, write_backward=0, namespaces=0, ksymbol=0, bpf_event=0, aux_output=0, cgroup=0, text_poke=0, build_id=0, inherit_thread=0, remove_on_exec=0, sigtrap=0, wakeup_events=0, config1=0, config2=0, sample_regs_user=0, sample_regs_intr=0, aux_watermark=0, sample_max_stack=0, aux_sample_size=0, sig_data=0}, -1, 0, -1, PERF_FLAG_FD_CLOEXEC) = -1 EINVAL (Invalid argument)
2162:perf_event_open({type=0x17 /* PERF_TYPE_??? */, size=PERF_ATTR_SIZE_VER7, config=0x2, sample_period=0, sample_type=PERF_SAMPLE_IDENTIFIER, read_format=PERF_FORMAT_TOTAL_TIME_ENABLED|PERF_FORMAT_TOTAL_TIME_RUNNING, disabled=1, inherit=1, pinned=0, exclusive=0, exclude_user=0, exclude_kernel=0, exclude_hv=0, exclude_idle=0, mmap=0, comm=0, freq=0, inherit_stat=0, enable_on_exec=0, task=0, watermark=0, precise_ip=0 /* arbitrary skid */, mmap_data=0, sample_id_all=0, exclude_host=0, exclude_guest=1, exclude_callchain_kernel=0, exclude_callchain_user=0, mmap2=0, comm_exec=0, use_clockid=0, context_switch=0, write_backward=0, namespaces=0, ksymbol=0, bpf_event=0, aux_output=0, cgroup=0, text_poke=0, build_id=0, inherit_thread=0, remove_on_exec=0, sigtrap=0, wakeup_events=0, config1=0, config2=0, sample_regs_user=0, sample_regs_intr=0, aux_watermark=0, sample_max_stack=0, aux_sample_size=0, sig_data=0}, -1, 0, -1, 0) = -1 EINVAL (Invalid argument)
2163:perf_event_open({type=0x17 /* PERF_TYPE_??? */, size=PERF_ATTR_SIZE_VER7, config=0x2, sample_period=0, sample_type=PERF_SAMPLE_IDENTIFIER, read_format=PERF_FORMAT_TOTAL_TIME_ENABLED|PERF_FORMAT_TOTAL_TIME_RUNNING, disabled=1, inherit=1, pinned=0, exclusive=0, exclude_user=0, exclude_kernel=0, exclude_hv=0, exclude_idle=0, mmap=0, comm=0, freq=0, inherit_stat=0, enable_on_exec=0, task=0, watermark=0, precise_ip=0 /* arbitrary skid */, mmap_data=0, sample_id_all=0, exclude_host=0, exclude_guest=0, exclude_callchain_kernel=0, exclude_callchain_user=0, mmap2=0, comm_exec=0, use_clockid=0, context_switch=0, write_backward=0, namespaces=0, ksymbol=0, bpf_event=0, aux_output=0, cgroup=0, text_poke=0, build_id=0, inherit_thread=0, remove_on_exec=0, sigtrap=0, wakeup_events=0, config1=0, config2=0, sample_regs_user=0, sample_regs_intr=0, aux_watermark=0, sample_max_stack=0, aux_sample_size=0, sig_data=0}, -1, 0, -1, 0) = 3
But how do we figure out what values to use for type and config? Scrolling up from the perf_event_open call I found a series of openat calls to various files. Here are some that matches the string energy-pkg
743:openat(AT_FDCWD, "/sys/bus/event_source/devices/power/events/energy-pkg", O_RDONLY) = 5
747:openat(AT_FDCWD, "/sys/bus/event_source/devices/power/events/energy-pkg.unit", O_RDONLY) = 6
750:openat(AT_FDCWD, "/sys/bus/event_source/devices/power/events/energy-pkg.scale", O_RDONLY) = 6
754:openat(AT_FDCWD, "/sys/bus/event_source/devices/power/events/energy-pkg.per-pkg", O_RDONLY) = -1 ENOENT (No such file or directory)
755:openat(AT_FDCWD, "/sys/bus/event_source/devices/power/events/energy-pkg.snapshot", O_RDONLY) = -1 ENOENT (No such file or directory)
So it looks like perf is reading those files to get the necessary configuration for the perf_event_open call. Here are the kernel docs for those files, that further validated my suspicion.
Here are the contents of those files on my box
$ sudo cat /sys/bus/event_source/devices/power/events/energy-pkg
event=0x02
$ sudo cat /sys/bus/event_source/devices/power/events/energy-pkg.unit
Joules
$ sudo cat /sys/bus/event_source/devices/power/events/energy-pkg.scale
2.3283064365386962890625e-10
Poking around in that directory, I found another file that looked important
$ sudo cat /sys/bus/event_source/devices/power/type
23
At this point, I was pretty confident I have all the info I need to use this syscall to measure power consumption. This is what I came up with
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>
#include <sys/types.h>
#include <sys/syscall.h>
#include <linux/perf_event.h>
#include <asm/unistd.h>
#define PERF_EVENT_OPEN_SYSCALL_NR __NR_perf_event_open
int main() {
int type = 0x17; // 23 in hex
long config = 0x02; // event code in hex
// Set up the perf_event_attr struct
struct perf_event_attr attr;
memset(&attr, 0, sizeof(attr));
attr.type = type;
attr.size = sizeof(struct perf_event_attr);
attr.config = config;
// Open the event
long fd = syscall(PERF_EVENT_OPEN_SYSCALL_NR, &attr, -1, 0, -1, 0);
if (fd == -1) {
perror("perf_event_open");
return 1;
}
// Read initial counter value
long int counter_before;
if (read(fd, &counter_before, sizeof(long int)) != sizeof(long int)) {
perror("read");
close(fd);
return 1;
}
printf("Measuring energy usage for 10 seconds...\n");
sleep(10);
// Read the final counter value
long int counter_after;
if (read(fd, &counter_after, sizeof(long int)) != sizeof(long int)) {
perror("read");
close(fd);
return 1;
}
// Calculate energy usage
double scale = 2.3283064365386962890625e-10; // from energy-cores.scale
double energy = (counter_after - counter_before) * scale;
printf("Energy usage: %.9f Joules\n", energy);
// Close the event
close(fd);
return 0;
}
Running this is simple
$ cc perf.c && sudo ./a.out
Measuring energy usage for 10 seconds...
Energy usage: 10.955932617 Joules
So the sleep(10) call consumes about 11 Joules in 10 seconds. That amounts to 1.1 Watts, since 1 Watt is 1 Joule per second.