Changelog v3-->v4: 1. Overhaul in implementation from kernel module to a userspace selftest ---
The patch series introduces a mechanism to measure wakeup latency for IPI and timer based interrupts The motivation behind this series is to find significant deviations behind advertised latency and residency values
To achieve this in the userspace, IPI latencies are calculated by sending information through pipes and inducing a wakeup, similarly alarm events are setup for calculate timer based wakeup latencies.
To account for delays from kernel-userspace interactions baseline observations are taken on a 100% busy CPU and subsequent obervations must be considered relative to that.
In theory, wakeups induced by IPI and Timers should have similar wakeup latencies, however in practice there may be deviations which may need to be captured.
One downside of the userspace approach in contrast to the kernel implementation is that the run to run variance can turn out to be high in the order of ms; which is the scope of the experiments at times.
Another downside of the userspace approach is that it takes much longer to run and hence a command-line option quick and full are added to make sure quick 1 CPU tests can be carried out when needed and otherwise it can carry out a full system comprehensive test.
Usage --- ./cpuidle --mode <full / quick / num_cpus> --output <output location> full: runs on all CPUS quick: run on a random CPU num_cpus: Limit the number of CPUS to run on
Sample output snippet --------------------- --IPI Latency Test--- SRC_CPU DEST_CPU IPI_Latency(ns) ... 0 5 256178 0 6 478161 0 7 285445 0 8 273553 Expected IPI latency(ns): 100000 Observed Average IPI latency(ns): 248334
--Timeout Latency Test-- --Baseline Timeout Latency measurement: CPU Busy-- Wakeup_src Baseline_delay(ns) ... 32 972405 33 1004287 34 986663 35 994022 Expected timeout(ns): 10000000 Observed Average timeout diff(ns): 991844
Pratik Rajesh Sampat (1): selftests/cpuidle: Add support for cpuidle latency measurement
tools/testing/selftests/Makefile | 1 + tools/testing/selftests/cpuidle/Makefile | 7 + tools/testing/selftests/cpuidle/cpuidle.c | 616 ++++++++++++++++++++++ tools/testing/selftests/cpuidle/settings | 1 + 4 files changed, 625 insertions(+) create mode 100644 tools/testing/selftests/cpuidle/Makefile create mode 100644 tools/testing/selftests/cpuidle/cpuidle.c create mode 100644 tools/testing/selftests/cpuidle/settings
Measure cpuidle latencies on wakeup to determine and compare with the advertsied wakeup latencies for each idle state.
Cpuidle wakeup latencies are determined for IPIs and Timer events and can help determine any deviations from what is advertsied by the hardware.
A baseline measurement for each case of IPI and timers is taken at 100 percent CPU usage to quantify for the kernel-userpsace overhead during execution.
Signed-off-by: Pratik Rajesh Sampat psampat@linux.ibm.com --- tools/testing/selftests/Makefile | 1 + tools/testing/selftests/cpuidle/Makefile | 7 + tools/testing/selftests/cpuidle/cpuidle.c | 616 ++++++++++++++++++++++ tools/testing/selftests/cpuidle/settings | 1 + 4 files changed, 625 insertions(+) create mode 100644 tools/testing/selftests/cpuidle/Makefile create mode 100644 tools/testing/selftests/cpuidle/cpuidle.c create mode 100644 tools/testing/selftests/cpuidle/settings
diff --git a/tools/testing/selftests/Makefile b/tools/testing/selftests/Makefile index 9018f45d631d..2bb0e87f76fd 100644 --- a/tools/testing/selftests/Makefile +++ b/tools/testing/selftests/Makefile @@ -8,6 +8,7 @@ TARGETS += cgroup TARGETS += clone3 TARGETS += core TARGETS += cpufreq +TARGETS += cpuidle TARGETS += cpu-hotplug TARGETS += drivers/dma-buf TARGETS += efivarfs diff --git a/tools/testing/selftests/cpuidle/Makefile b/tools/testing/selftests/cpuidle/Makefile new file mode 100644 index 000000000000..d332485e1bc5 --- /dev/null +++ b/tools/testing/selftests/cpuidle/Makefile @@ -0,0 +1,7 @@ +# SPDX-License-Identifier: GPL-2.0 +TEST_GEN_PROGS := cpuidle + +CFLAGS += -O2 +LDLIBS += -lpthread + +include ../lib.mk diff --git a/tools/testing/selftests/cpuidle/cpuidle.c b/tools/testing/selftests/cpuidle/cpuidle.c new file mode 100644 index 000000000000..4b1e7a91f75c --- /dev/null +++ b/tools/testing/selftests/cpuidle/cpuidle.c @@ -0,0 +1,616 @@ +// SPDX-License-Identifier: GPL-2.0-or-later +/* + * Cpuidle latency measurement microbenchmark + * + * A mechanism to measure wakeup latency for IPI and Timer based interrupts + * Results of this microbenchmark can be used to check and validate against the + * advertised latencies for each cpuidle state + * + * IPIs (using pipes) and Timers are used to wake the CPU up and measure the + * time difference + * + * Usage: + * ./cpuidle --mode <full / quick / num_cpus> --output <output location> + * + * Copyright (C) 2020 Pratik Rajesh Sampat psampat@linux.ibm.com, IBM + */ + +#define _GNU_SOURCE +#include <assert.h> +#include <dirent.h> +#include <fcntl.h> +#include <getopt.h> +#include <inttypes.h> +#include <limits.h> +#include <pthread.h> +#include <sched.h> +#include <signal.h> +#include <stdio.h> +#include <stdlib.h> +#include <string.h> +#include <sys/time.h> +#include <unistd.h> + +#define READ 0 +#define WRITE 1 +#define TIMEOUT_US 500000 + +static int pipe_fd[2]; +static int *cpu_list; +static int cpus; +static int idle_states; +static uint64_t *latency_list; +static uint64_t *residency_list; + +static char *log_file = "cpuidle.log"; + +static int get_online_cpus(int *online_cpu_list, int total_cpus) +{ + char filename[80]; + int i, index = 0; + FILE *fptr; + + for (i = 0; i < total_cpus; i++) { + char status; + + sprintf(filename, "/sys/devices/system/cpu/cpu"); + sprintf(filename + strlen(filename), "%d%s", i, "/online"); + fptr = fopen(filename, "r"); + if (!fptr) + continue; + assert(fscanf(fptr, "%c", &status) != EOF); + if (status == '1') + online_cpu_list[index++] = i; + fclose(fptr); + } + return index; +} + +static uint64_t us_to_ns(uint64_t val) +{ + return val * 1000; +} + +static void get_latency(int cpu) +{ + char filename[80]; + uint64_t latency; + FILE *fptr; + int state; + + for (state = 0; state < idle_states; state++) { + sprintf(filename, "%s%d%s%d%s", "/sys/devices/system/cpu/cpu", + cpu, "/cpuidle/state", + state, "/latency"); + fptr = fopen(filename, "r"); + assert(fptr); + + assert(fscanf(fptr, "%ld", &latency) != EOF); + latency_list[state] = latency; + fclose(fptr); + } +} + +static void get_residency(int cpu) +{ + uint64_t residency; + char filename[80]; + FILE *fptr; + int state; + + for (state = 0; state < idle_states; state++) { + sprintf(filename, "%s%d%s%d%s", "/sys/devices/system/cpu/cpu", + cpu, "/cpuidle/state", + state, "/residency"); + fptr = fopen(filename, "r"); + assert(fptr); + + assert(fscanf(fptr, "%ld", &residency) != EOF); + residency_list[state] = residency; + fclose(fptr); + } +} + +static int get_idle_state_count(int cpu) +{ + struct dirent *entry; + int dir_count = 0; + char filename[80]; + DIR *dirp; + + sprintf(filename, "%s%d%s", "/sys/devices/system/cpu/cpu", + cpu, "/cpuidle"); + + dirp = opendir(filename); + if (!dirp) + return -1; + while (entry = readdir(dirp)) { + if (entry->d_type == DT_DIR) { + if (strcmp(entry->d_name, ".") == 0 || + strcmp(entry->d_name, "..") == 0) + continue; + dir_count++; + } + } + closedir(dirp); + return dir_count; +} + +/* Enable or disable all idle states */ +static int state_all_idle(char *disable) +{ + char filename[80]; + FILE *fptr; + int i, j; + + for (i = 0; i < cpus; i++) { + for (j = 0; j < idle_states; j++) { + sprintf(filename, "%s%d%s%d%s", + "/sys/devices/system/cpu/cpu", cpu_list[i], + "/cpuidle/state", j, "/disable"); + fptr = fopen(filename, "w"); + assert(fptr); + fprintf(fptr, "%s", disable); + fclose(fptr); + } + } + return 0; +} + +/* Disable all idle states */ +static int cpuidle_disable_all_states(void) +{ + return state_all_idle("1"); +} + +static int cpuidle_enable_all_states(void) +{ + return state_all_idle("0"); +} + +static int state_operation(char *disable, int state) +{ + char filename[80]; + FILE *fptr; + int i; + + for (i = 0; i < cpus; i++) { + sprintf(filename, "%s%d%s%d%s", "/sys/devices/system/cpu/cpu", + cpu_list[i], "/cpuidle/state", state, "/disable"); + fptr = fopen(filename, "w"); + assert(fptr); + fprintf(fptr, "%s", disable); + fclose(fptr); + } + return 0; +} + +static int cpuidle_enable_state(int state) +{ + return state_operation("0", state); +} + +static int cpuidle_disable_state(int state) +{ + return state_operation("1", state); +} + +static uint64_t average(uint64_t *arr, int size) +{ + int i, sum = 0; + + assert(size != 0); + for (i = 0; i < size; i++) + sum += arr[i]; + return sum / size; +} + +static pthread_t start_thread_on(void *(*fn)(void *), void *arg, uint64_t cpu) +{ + pthread_attr_t attr; + cpu_set_t cpuset; + pthread_t tid; + int rc; + + CPU_ZERO(&cpuset); + CPU_SET(cpu, &cpuset); + + rc = pthread_attr_init(&attr); + if (rc) { + perror("pthread_attr_init"); + exit(1); + } + + rc = pthread_attr_setaffinity_np(&attr, sizeof(cpu_set_t), &cpuset); + if (rc) { + perror("pthread_attr_setaffinity_np"); + exit(1); + } + + rc = pthread_create(&tid, &attr, fn, arg); + if (rc) { + perror("pthread_create"); + exit(1); + } + return tid; +} + +void *util_full_cpu(void *unused) +{ + FILE *fptr; + + fptr = fopen("/dev/null", "w"); + assert(fptr); + while (1) { + fprintf(fptr, "0"); + nanosleep((const struct timespec[]){{0, 10L}}, NULL); + } + fclose(fptr); +} + +/* IPI based wakeup latencies */ +struct latency { + unsigned int src_cpu; + unsigned int dest_cpu; + struct timespec time_start; + struct timespec time_end; + uint64_t latency_ns; +} ipi_wakeup; + +static void *writer(void *unused) +{ + signed char c = 'P'; + + assert(write(pipe_fd[WRITE], &c, 1) == 1); + ipi_wakeup.src_cpu = sched_getcpu(); + + return NULL; +} + +static void *reader(void *unused) +{ + signed char c; + + assert(read(pipe_fd[READ], &c, 1) == 1); + ipi_wakeup.dest_cpu = sched_getcpu(); + + return NULL; +} + +static void ipi_test_once(int baseline, int src_cpu, int dest_cpu) +{ + pthread_t tid, tid1, baseline_tid; + + if (baseline) { + baseline_tid = start_thread_on(util_full_cpu, NULL, dest_cpu); + /* Run process for long enough to gain 100% usage*/ + sleep(2); + } + + clock_gettime(CLOCK_REALTIME, &ipi_wakeup.time_start); + + tid = start_thread_on(writer, NULL, src_cpu); + pthread_join(tid, NULL); + tid1 = start_thread_on(reader, NULL, dest_cpu); + pthread_join(tid1, NULL); + + clock_gettime(CLOCK_REALTIME, &ipi_wakeup.time_end); + ipi_wakeup.latency_ns = (ipi_wakeup.time_end.tv_sec - + ipi_wakeup.time_start.tv_sec) * 1000000000ULL + + ipi_wakeup.time_end.tv_nsec - ipi_wakeup.time_start.tv_nsec; + + if (baseline) + pthread_cancel(baseline_tid); +} + +static void ipi_test(int src_cpu) +{ + uint64_t avg_arr[cpus], avg_latency; + int cpu, state; + FILE *fptr; + + assert(cpuidle_disable_all_states() == 0); + + if (pipe(pipe_fd)) + exit(1); + + fptr = fopen(log_file, "a"); + fprintf(fptr, "----IPI TEST----\n"); + + fprintf(fptr, "----Baseline IPI Latency----\n"); + fprintf(fptr, "%s %10s %18s\n", "SRC_CPU", "DEST_CPU", + "Baseline_latency(ns)"); + /* Run the test as dummy once to stablize */ + ipi_test_once(1, src_cpu, cpu_list[0]); + for (cpu = 0; cpu < cpus; cpu++) { + ipi_test_once(1, src_cpu, cpu_list[cpu]); + fprintf(fptr, "%3d %10d %12ld\n", ipi_wakeup.src_cpu, + ipi_wakeup.dest_cpu, + ipi_wakeup.latency_ns); + avg_arr[cpu] = ipi_wakeup.latency_ns; + } + avg_latency = average(avg_arr, cpus); + fprintf(fptr, "Baseline average IPI latency(ns): %ld\n\n", avg_latency); + + for (state = 0; state < idle_states; state++) { + fprintf(fptr, "--Enabling state: %d--\n", state); + assert(cpuidle_enable_state(state) == 0); + fprintf(fptr, "%s %10s %18s\n", "SRC_CPU", "DEST_CPU", + "IPI_Latency(ns)"); + for (cpu = 0; cpu < cpus; cpu++) { + /* Allow sufficient cycles to go idle */ + sleep(1); + ipi_test_once(0, src_cpu, cpu_list[cpu]); + fprintf(fptr, "%3d %10d %18ld\n", ipi_wakeup.src_cpu, + ipi_wakeup.dest_cpu, + ipi_wakeup.latency_ns); + avg_arr[cpu] = ipi_wakeup.latency_ns; + } + fprintf(fptr, "Expected Latency(ns): %ld\n", + us_to_ns(latency_list[state])); + avg_latency = average(avg_arr, cpus); + fprintf(fptr, "Observed Average IPI latency(ns): %ld\n\n", + avg_latency); + assert(cpuidle_disable_state(state) == 0); + } + + assert(cpuidle_enable_all_states() == 0); + fclose(fptr); +} + +/* Timer based wakeup latencies */ +static int soak_done; +struct timer_data { + unsigned int src_cpu; + uint64_t timeout; + struct timespec time_start; + struct timespec time_end; + uint64_t timeout_diff_ns; +} timer_wakeup; + +void catch_alarm(int sig) +{ + soak_done = 1; +} + +static void setup_timer(void) +{ + struct itimerval timer_settings = {}; + int err; + + timer_settings.it_value.tv_usec = timer_wakeup.timeout; + err = setitimer(ITIMER_REAL, &timer_settings, NULL); + if (err < 0) { + perror("failed to arm interval timer"); + exit(1); + } + signal(SIGALRM, catch_alarm); + while (!soak_done) + sleep(1); +} + +static void *queue_timer(void *timeout) +{ + timer_wakeup.src_cpu = sched_getcpu(); + timer_wakeup.timeout = (uint64_t)timeout; + setup_timer(); + + return NULL; +} + +static void timeout_test_once(int baseline, uint64_t timeout, int dest_cpu) +{ + pthread_t tid, baseline_tid; + + if (baseline) { + baseline_tid = start_thread_on(util_full_cpu, NULL, dest_cpu); + /* Run process for long enough to gain 100% usage */ + sleep(2); + } + + clock_gettime(CLOCK_REALTIME, &timer_wakeup.time_start); + + tid = start_thread_on(queue_timer, (void *)timeout, dest_cpu); + pthread_join(tid, NULL); + + clock_gettime(CLOCK_REALTIME, &timer_wakeup.time_end); + timer_wakeup.timeout_diff_ns = (timer_wakeup.time_end.tv_sec - + timer_wakeup.time_start.tv_sec) * 1000000000ULL + + (timer_wakeup.time_end.tv_nsec - timer_wakeup.time_start.tv_nsec); + if (baseline) + pthread_cancel(baseline_tid); +} + +static void timeout_test(unsigned long timeout) +{ + uint64_t avg_arr[cpus], avg_timeout_diff; + int state, cpu; + FILE *fptr; + + assert(cpuidle_disable_all_states() == 0); + fptr = fopen(log_file, "a"); + fprintf(fptr, "----TIMEOUT TEST----\n"); + + fprintf(fptr, "----Baseline Timeout Latency diff----\n"); + fprintf(fptr, "%s %10s\n", "SRC_CPU", "Baseline_Latency(ns)"); + /* Run the test as dummy once to stablize */ + timeout_test_once(1, timeout, cpu_list[0]); + for (cpu = 0; cpu < cpus; cpu++) { + timeout_test_once(1, timeout, cpu_list[cpu]); + fprintf(fptr, "%3d %11ld\n", timer_wakeup.src_cpu, + timer_wakeup.timeout_diff_ns); + avg_arr[cpu] = timer_wakeup.timeout_diff_ns; + } + avg_timeout_diff = average(avg_arr, cpus); + fprintf(fptr, "Baseline Average Timeout diff(ns): %ld\n\n", + avg_timeout_diff); + + for (state = 0; state < idle_states; state++) { + fprintf(fptr, "--Enabling state: %d--\n", state); + assert(cpuidle_enable_state(state) == 0); + fprintf(fptr, "%s %11s\n", "SRC_CPU", "Timeout_Latency(ns)"); + for (cpu = 0; cpu < cpus; cpu++) { + /* Allow sufficient cycles to go idle */ + sleep(1); + timeout_test_once(0, timeout, + cpu_list[cpu]); + fprintf(fptr, "%3d %15ld\n", timer_wakeup.src_cpu, + timer_wakeup.timeout_diff_ns); + avg_arr[cpu] = timer_wakeup.timeout_diff_ns; + } + fprintf(fptr, "Expected Residency(ns): %ld\n", + us_to_ns(residency_list[state])); + avg_timeout_diff = average(avg_arr, cpus); + fprintf(fptr, "Observed Average Timeout diff(ns): %ld\n\n", + avg_timeout_diff); + assert(cpuidle_disable_state(state) == 0); + } + assert(cpuidle_enable_all_states() == 0); + fclose(fptr); +} + +static struct option options[] = { + {"output", required_argument, 0, 'o'}, + {"mode", required_argument, 0, 'm'}, +}; + +static void usage(void) +{ + fprintf(stderr, "Usage: cpuidle <options>\n\n"); + fprintf(stderr, "\t\t--mode=X\t(quick / full / <num_cpus>)\n"); + fprintf(stderr, "\t\t--output=X\tOutput loc (def: cpuidle.log)\n\n"); +} + +void get_n_random_cpus(int *shf_list, int shf_size, int *list, int n) +{ + /* Shuffle the Online CPU list */ + int i; + int shuffle_index_list[shf_size]; + + for (i = 0; i < shf_size; i++) + shuffle_index_list[i] = i; + for (i = 0; i < shf_size; i++) { + int idx = i + rand() % (shf_size - i); + int temp = shuffle_index_list[i]; + + shuffle_index_list[i] = shuffle_index_list[idx]; + shuffle_index_list[idx] = temp; + } + + /* Pick the first n from the shf_list elements */ + for (i = 0; i < n; i++) + list[i] = shf_list[shuffle_index_list[i]]; +} + +int main(int argc, char *argv[]) +{ + int total_cpus, online_cpus, option_index = 0; + int *online_cpu_list; + signed char c; + FILE *fptr; + + if (getuid()) { + fprintf(stderr, "cpuidle latency test must run as root\n"); + exit(1); + } + + total_cpus = sysconf(_SC_NPROCESSORS_ONLN); + online_cpu_list = malloc(total_cpus * sizeof(int)); + if (!online_cpu_list) { + perror("Malloc failure"); + exit(1); + } + + online_cpus = get_online_cpus(&online_cpu_list[0], total_cpus); + if (!online_cpus) { + perror("Unable to get online CPUS"); + exit(1); + } + + /* Get one CPU for a quick test */ + cpus = 1; + cpu_list = malloc(1 * sizeof(int)); + srand(time(NULL)); + cpu_list[0] = online_cpu_list[(rand() % online_cpus) + 1]; + + while (1) { + c = getopt_long(argc, argv, "", options, &option_index); + if (c == -1) + break; + + switch (c) { + case 'o': + log_file = optarg; + break; + case 'm': + if (!strcmp(optarg, "full")) { + cpu_list = realloc(cpu_list, + online_cpus * sizeof(int)); + memcpy(cpu_list, online_cpu_list, + online_cpus * sizeof(int)); + cpus = online_cpus; + } else if (strcmp(optarg, "quick")) { + int opt_cpus; + + opt_cpus = atoi(optarg); + if (!opt_cpus) { + fprintf(stderr, "Error parsing mode\n"); + usage(); + exit(1); + } + if (opt_cpus > online_cpus) { + fprintf(stderr, "Number of CPUS > Online CPUs\n"); + usage(); + exit(1); + } + cpu_list = realloc(cpu_list, + opt_cpus * sizeof(int)); + get_n_random_cpus(online_cpu_list, online_cpus, + &cpu_list[0], opt_cpus); + cpus = opt_cpus; + } + break; + default: + usage(); + exit(1); + } + } + + idle_states = get_idle_state_count(online_cpu_list[0]); + if (idle_states == -1) { + perror("Unable to get idle states"); + exit(1); + } + + fptr = fopen(log_file, "w+"); + fprintf(fptr, "cpuidle latency selftests. IPI & Timers\n"); + fprintf(fptr, "Number of CPUS: %d\n", total_cpus); + fprintf(fptr, "Number of idle states: %d\n", idle_states); + fclose(fptr); + + latency_list = malloc(idle_states * sizeof(uint64_t)); + residency_list = malloc(idle_states * sizeof(uint64_t)); + if (!latency_list || !residency_list) { + perror("Malloc failure"); + exit(1); + } + + get_latency(online_cpu_list[0]); + get_residency(online_cpu_list[0]); + + ipi_test(online_cpu_list[0]); + printf("IPI test done\n"); + fflush(stdout); + + timeout_test(TIMEOUT_US); + printf("Timeout test done\n"); + fflush(stdout); + + printf("Output logged at: %s\n", log_file); + + free(latency_list); + free(residency_list); + free(cpu_list); + free(online_cpu_list); + return 0; +} diff --git a/tools/testing/selftests/cpuidle/settings b/tools/testing/selftests/cpuidle/settings new file mode 100644 index 000000000000..e7b9417537fb --- /dev/null +++ b/tools/testing/selftests/cpuidle/settings @@ -0,0 +1 @@ +timeout=0
On Wed, 2020-09-02 at 17:15 +0530, Pratik Rajesh Sampat wrote:
Measure cpuidle latencies on wakeup to determine and compare with the advertsied wakeup latencies for each idle state.
It looks like the measurements include more than just C-state wake, they also include the overhead of waking up the proces, context switch, and potentially any interrupts that happen on that CPU. I am not saying this is not interesting data, it surely is, but it is going to be larger than you see in cpuidle latency tables. Potentially significantly larger.
Therefore, I am not sure this program should be advertised as "cpuidle measurement". It really measures the "IPI latency" in case of the IPI method.
A baseline measurement for each case of IPI and timers is taken at 100 percent CPU usage to quantify for the kernel-userpsace overhead during execution.
At least on Intel platforms, this will mean that the IPI method won't cover deep C-states like, say, PC6, because one CPU is busy. Again, not saying this is not interesting, just pointing out the limitation.
I was working on a somewhat similar stuff for x86 platforms, and I am almost ready to publish that on github. I can notify you when I do so if you are interested. But here is a small presentation of the approach that I did on Plumbers last year:
https://youtu.be/Opk92aQyvt0?t=8266
(the link points to the start of my talk)
Hello Artem,
On 02/09/20 8:55 pm, Artem Bityutskiy wrote:
On Wed, 2020-09-02 at 17:15 +0530, Pratik Rajesh Sampat wrote:
Measure cpuidle latencies on wakeup to determine and compare with the advertsied wakeup latencies for each idle state.
Thank you for pointing me to your talk. It was very interesting! I certainly did not know about that the Intel architecture being aware of timers and pre-wakes the CPUs which makes the timer experiment observations void.
It looks like the measurements include more than just C-state wake, they also include the overhead of waking up the proces, context switch, and potentially any interrupts that happen on that CPU. I am not saying this is not interesting data, it surely is, but it is going to be larger than you see in cpuidle latency tables. Potentially significantly larger.
The measurements will definitely include overhead than just the C-State wakeup.
However, we are also collecting a baseline measurement wherein we run the same test on a 100% busy CPU and the measurement of latency from that could be considered to the kernel-userspace overhead. The rest of the measurements would be considered keeping this baseline in mind.
Therefore, I am not sure this program should be advertised as "cpuidle measurement". It really measures the "IPI latency" in case of the IPI method.
Now with the new found knowledge of timers in Intel, I understand that this really only seems to measure IPI latency and not timer latency, although both the observations shouldn't be too far off anyways.
A baseline measurement for each case of IPI and timers is taken at 100 percent CPU usage to quantify for the kernel-userpsace overhead during execution.
At least on Intel platforms, this will mean that the IPI method won't cover deep C-states like, say, PC6, because one CPU is busy. Again, not saying this is not interesting, just pointing out the limitation.
That's a valid point. We have similar deep idle states in POWER too. The idea here is that this test should be run on an already idle system, of course there will be kernel jitters along the way which can cause little skewness in observations across some CPUs but I believe the observations overall should be stable.
Another solution to this could be using isolcpus, but that just increases the complexity all the more. If you have any suggestions of any other way that could guarantee idleness that would be great.
I was working on a somewhat similar stuff for x86 platforms, and I am almost ready to publish that on github. I can notify you when I do so if you are interested. But here is a small presentation of the approach that I did on Plumbers last year:
https://youtu.be/Opk92aQyvt0?t=8266
(the link points to the start of my talk)
Sure thing. Do notify me when it comes up. I would be happy to have a look at it.
-- Thanks! Pratik
On Thu, 2020-09-03 at 17:30 +0530, Pratik Sampat wrote:
I certainly did not know about that the Intel architecture being aware of timers and pre-wakes the CPUs which makes the timer experiment observations void.
Well, things depend on platform, it is really "void", it is just different and it measures an optimized case. The result may be smaller observed latency. And things depend on the platform.
However, we are also collecting a baseline measurement wherein we run the same test on a 100% busy CPU and the measurement of latency from that could be considered to the kernel-userspace overhead. The rest of the measurements would be considered keeping this baseline in mind.
Yes, this should give the idea of the overhead, but still, at least for many Intel platforms I would not be comfortable using the resulting number (measured latency - baseline) for a cpuidle driver, because there are just too many variables there. I am not sure I could assume the baseline measured this way is an invariant - it could be noticeably different depending on whether you use C-states or not.
At least on Intel platforms, this will mean that the IPI method won't cover deep C-states like, say, PC6, because one CPU is busy. Again, not saying this is not interesting, just pointing out the limitation.
That's a valid point. We have similar deep idle states in POWER too. The idea here is that this test should be run on an already idle system, of course there will be kernel jitters along the way which can cause little skewness in observations across some CPUs but I believe the observations overall should be stable.
If baseline and cpuidle latency are numbers of same order of magnitude, and you are measuring in a controlled lab system, may be yes. But if baseline is, say, in milliseconds, and you are measuring a 10 microseconds C-state, then probably no.
Another solution to this could be using isolcpus, but that just increases the complexity all the more. If you have any suggestions of any other way that could guarantee idleness that would be great.
Well, I did not try to guarantee idleness. I just use timers and external device (the network card), so no CPUs needs to be busy and the system can enter deep C-states. Then I just look at median, 99%-th percentile, etc.
But by all means IPI is also a very interesting experiment. Just covers a different usage scenario.
When I started experimenting in this area, one of my main early takeaways was realization that C-state latency really depends on the event source.
HTH, Artem.
On Thu, 2020-09-03 at 17:50 +0300, Artem Bityutskiy wrote:
Well, things depend on platform, it is really "void", it is just different and it measures an optimized case. The result may be smaller observed latency.
Sorry, I meant to say it is _not_ really "void".
On 03/09/20 8:20 pm, Artem Bityutskiy wrote:
On Thu, 2020-09-03 at 17:30 +0530, Pratik Sampat wrote:
I certainly did not know about that the Intel architecture being aware of timers and pre-wakes the CPUs which makes the timer experiment observations void.
Well, things depend on platform, it is really "void", it is just different and it measures an optimized case. The result may be smaller observed latency. And things depend on the platform.
Of course, this will be for just software observability and hardware can be more complex with each architecture behaving differently.
However, we are also collecting a baseline measurement wherein we run the same test on a 100% busy CPU and the measurement of latency from that could be considered to the kernel-userspace overhead. The rest of the measurements would be considered keeping this baseline in mind.
Yes, this should give the idea of the overhead, but still, at least for many Intel platforms I would not be comfortable using the resulting number (measured latency - baseline) for a cpuidle driver, because there are just too many variables there. I am not sure I could assume the baseline measured this way is an invariant - it could be noticeably different depending on whether you use C-states or not.
At least on Intel platforms, this will mean that the IPI method won't cover deep C-states like, say, PC6, because one CPU is busy. Again, not saying this is not interesting, just pointing out the limitation.
That's a valid point. We have similar deep idle states in POWER too. The idea here is that this test should be run on an already idle system, of course there will be kernel jitters along the way which can cause little skewness in observations across some CPUs but I believe the observations overall should be stable.
If baseline and cpuidle latency are numbers of same order of magnitude, and you are measuring in a controlled lab system, may be yes. But if baseline is, say, in milliseconds, and you are measuring a 10 microseconds C-state, then probably no.
This makes complete sense. The magnitude of deviations being greater than the scope of the experiment may not be very useful in quantifying the latency metric.
One way is to minimize the baseline overhead is to make this a kernel module https://lkml.org/lkml/2020/7/21/567. However, the overhead is unavoidable but definetly can be further minimized by using an external approach suggested by you in your LPC talk
Another solution to this could be using isolcpus, but that just increases the complexity all the more. If you have any suggestions of any other way that could guarantee idleness that would be great.
Well, I did not try to guarantee idleness. I just use timers and external device (the network card), so no CPUs needs to be busy and the system can enter deep C-states. Then I just look at median, 99%-th percentile, etc.
But by all means IPI is also a very interesting experiment. Just covers a different usage scenario.
When I started experimenting in this area, one of my main early takeaways was realization that C-state latency really depends on the event source.
That is an interesting observation, on POWER systems where we don't have timer related wakeup optimizations, the readings from this test do signify a difference between latencies of IPI versus the latency gathered after a timer interrupt.
However, these timer based variations weren't as prominent on my Intel based ThinkPad t480, therefore in confirmation with your observations.
This discussions does help! Although this approach may not help quantify latency deviations at a hardware-accurate level but could still be helpful in quantifying this metric from a software observability point of view.
Thanks! Pratik
On Wed, Sep 02, 2020 at 05:15:06PM +0530, Pratik Rajesh Sampat wrote:
Measure cpuidle latencies on wakeup to determine and compare with the advertsied wakeup latencies for each idle state.
Cpuidle wakeup latencies are determined for IPIs and Timer events and can help determine any deviations from what is advertsied by the hardware.
A baseline measurement for each case of IPI and timers is taken at 100 percent CPU usage to quantify for the kernel-userpsace overhead during execution.
Signed-off-by: Pratik Rajesh Sampat psampat@linux.ibm.com
tools/testing/selftests/Makefile | 1 + tools/testing/selftests/cpuidle/Makefile | 7 + tools/testing/selftests/cpuidle/cpuidle.c | 616 ++++++++++++++++++++++ tools/testing/selftests/cpuidle/settings | 1 + 4 files changed, 625 insertions(+) create mode 100644 tools/testing/selftests/cpuidle/Makefile create mode 100644 tools/testing/selftests/cpuidle/cpuidle.c create mode 100644 tools/testing/selftests/cpuidle/settings
diff --git a/tools/testing/selftests/Makefile b/tools/testing/selftests/Makefile index 9018f45d631d..2bb0e87f76fd 100644 --- a/tools/testing/selftests/Makefile +++ b/tools/testing/selftests/Makefile @@ -8,6 +8,7 @@ TARGETS += cgroup TARGETS += clone3 TARGETS += core TARGETS += cpufreq +TARGETS += cpuidle TARGETS += cpu-hotplug TARGETS += drivers/dma-buf TARGETS += efivarfs diff --git a/tools/testing/selftests/cpuidle/Makefile b/tools/testing/selftests/cpuidle/Makefile new file mode 100644 index 000000000000..d332485e1bc5 --- /dev/null +++ b/tools/testing/selftests/cpuidle/Makefile @@ -0,0 +1,7 @@ +# SPDX-License-Identifier: GPL-2.0 +TEST_GEN_PROGS := cpuidle
+CFLAGS += -O2 +LDLIBS += -lpthread
+include ../lib.mk diff --git a/tools/testing/selftests/cpuidle/cpuidle.c b/tools/testing/selftests/cpuidle/cpuidle.c new file mode 100644 index 000000000000..4b1e7a91f75c --- /dev/null +++ b/tools/testing/selftests/cpuidle/cpuidle.c @@ -0,0 +1,616 @@ +// SPDX-License-Identifier: GPL-2.0-or-later +/*
- Cpuidle latency measurement microbenchmark
- A mechanism to measure wakeup latency for IPI and Timer based interrupts
- Results of this microbenchmark can be used to check and validate against the
- advertised latencies for each cpuidle state
- IPIs (using pipes) and Timers are used to wake the CPU up and measure the
- time difference
- Usage:
- ./cpuidle --mode <full / quick / num_cpus> --output <output location>
- Copyright (C) 2020 Pratik Rajesh Sampat psampat@linux.ibm.com, IBM
- */
+#define _GNU_SOURCE +#include <assert.h> +#include <dirent.h> +#include <fcntl.h> +#include <getopt.h> +#include <inttypes.h> +#include <limits.h> +#include <pthread.h> +#include <sched.h> +#include <signal.h> +#include <stdio.h> +#include <stdlib.h> +#include <string.h> +#include <sys/time.h> +#include <unistd.h>
+#define READ 0 +#define WRITE 1 +#define TIMEOUT_US 500000
+static int pipe_fd[2]; +static int *cpu_list; +static int cpus; +static int idle_states; +static uint64_t *latency_list; +static uint64_t *residency_list;
+static char *log_file = "cpuidle.log";
+static int get_online_cpus(int *online_cpu_list, int total_cpus) +{
- char filename[80];
- int i, index = 0;
- FILE *fptr;
- for (i = 0; i < total_cpus; i++) {
char status;sprintf(filename, "/sys/devices/system/cpu/cpu");sprintf(filename + strlen(filename), "%d%s", i, "/online");fptr = fopen(filename, "r");if (!fptr)continue;assert(fscanf(fptr, "%c", &status) != EOF);if (status == '1')online_cpu_list[index++] = i;fclose(fptr);- }
- return index;
+}
+static uint64_t us_to_ns(uint64_t val) +{
- return val * 1000;
+}
+static void get_latency(int cpu) +{
- char filename[80];
- uint64_t latency;
- FILE *fptr;
- int state;
- for (state = 0; state < idle_states; state++) {
sprintf(filename, "%s%d%s%d%s", "/sys/devices/system/cpu/cpu",cpu, "/cpuidle/state",state, "/latency");fptr = fopen(filename, "r");assert(fptr);assert(fscanf(fptr, "%ld", &latency) != EOF);latency_list[state] = latency;fclose(fptr);- }
+}
+static void get_residency(int cpu) +{
- uint64_t residency;
- char filename[80];
- FILE *fptr;
- int state;
- for (state = 0; state < idle_states; state++) {
sprintf(filename, "%s%d%s%d%s", "/sys/devices/system/cpu/cpu",cpu, "/cpuidle/state",state, "/residency");fptr = fopen(filename, "r");assert(fptr);assert(fscanf(fptr, "%ld", &residency) != EOF);residency_list[state] = residency;fclose(fptr);- }
+}
+static int get_idle_state_count(int cpu) +{
- struct dirent *entry;
- int dir_count = 0;
- char filename[80];
- DIR *dirp;
- sprintf(filename, "%s%d%s", "/sys/devices/system/cpu/cpu",
cpu, "/cpuidle");- dirp = opendir(filename);
- if (!dirp)
return -1;- while (entry = readdir(dirp)) {
if (entry->d_type == DT_DIR) {if (strcmp(entry->d_name, ".") == 0 ||strcmp(entry->d_name, "..") == 0)continue;dir_count++;}- }
- closedir(dirp);
- return dir_count;
+}
+/* Enable or disable all idle states */ +static int state_all_idle(char *disable) +{
- char filename[80];
- FILE *fptr;
- int i, j;
- for (i = 0; i < cpus; i++) {
for (j = 0; j < idle_states; j++) {sprintf(filename, "%s%d%s%d%s","/sys/devices/system/cpu/cpu", cpu_list[i],"/cpuidle/state", j, "/disable");fptr = fopen(filename, "w");assert(fptr);fprintf(fptr, "%s", disable);fclose(fptr);}- }
- return 0;
+}
+/* Disable all idle states */ +static int cpuidle_disable_all_states(void) +{
- return state_all_idle("1");
+}
+static int cpuidle_enable_all_states(void) +{
- return state_all_idle("0");
+}
+static int state_operation(char *disable, int state) +{
- char filename[80];
- FILE *fptr;
- int i;
- for (i = 0; i < cpus; i++) {
sprintf(filename, "%s%d%s%d%s", "/sys/devices/system/cpu/cpu",cpu_list[i], "/cpuidle/state", state, "/disable");fptr = fopen(filename, "w");assert(fptr);fprintf(fptr, "%s", disable);fclose(fptr);- }
- return 0;
+}
+static int cpuidle_enable_state(int state) +{
- return state_operation("0", state);
+}
+static int cpuidle_disable_state(int state) +{
- return state_operation("1", state);
+}
+static uint64_t average(uint64_t *arr, int size) +{
- int i, sum = 0;
- assert(size != 0);
- for (i = 0; i < size; i++)
sum += arr[i];- return sum / size;
+}
+static pthread_t start_thread_on(void *(*fn)(void *), void *arg, uint64_t cpu) +{
- pthread_attr_t attr;
- cpu_set_t cpuset;
- pthread_t tid;
- int rc;
- CPU_ZERO(&cpuset);
- CPU_SET(cpu, &cpuset);
- rc = pthread_attr_init(&attr);
- if (rc) {
perror("pthread_attr_init");exit(1);- }
- rc = pthread_attr_setaffinity_np(&attr, sizeof(cpu_set_t), &cpuset);
- if (rc) {
perror("pthread_attr_setaffinity_np");exit(1);- }
- rc = pthread_create(&tid, &attr, fn, arg);
- if (rc) {
perror("pthread_create");exit(1);- }
- return tid;
+}
+void *util_full_cpu(void *unused) +{
- FILE *fptr;
- fptr = fopen("/dev/null", "w");
- assert(fptr);
- while (1) {
fprintf(fptr, "0");nanosleep((const struct timespec[]){{0, 10L}}, NULL);
It is not clear why the nanosleep is needed. Perhaps you want the equivalent of cpu_relax() in the linux kernel ?
- }
- fclose(fptr);
+}
+/* IPI based wakeup latencies */ +struct latency {
- unsigned int src_cpu;
- unsigned int dest_cpu;
- struct timespec time_start;
- struct timespec time_end;
- uint64_t latency_ns;
+} ipi_wakeup;
+static void *writer(void *unused) +{
- signed char c = 'P';
- assert(write(pipe_fd[WRITE], &c, 1) == 1);
- ipi_wakeup.src_cpu = sched_getcpu();
- return NULL;
+}
+static void *reader(void *unused) +{
- signed char c;
- assert(read(pipe_fd[READ], &c, 1) == 1);
- ipi_wakeup.dest_cpu = sched_getcpu();
- return NULL;
+}
+static void ipi_test_once(int baseline, int src_cpu, int dest_cpu) +{
- pthread_t tid, tid1, baseline_tid;
- if (baseline) {
baseline_tid = start_thread_on(util_full_cpu, NULL, dest_cpu);/* Run process for long enough to gain 100% usage*/sleep(2);- }
- clock_gettime(CLOCK_REALTIME, &ipi_wakeup.time_start);
If we capture the time_start here, it would also include the time required to create a thread and run it on src_cpu.
Shouldn't we move clock_gettime(time_start) inside the writer, just before the write(pipe) and another clock_gettime(time_end) in the reader soon after the read(pipe) returns ?
- tid = start_thread_on(writer, NULL, src_cpu);
- pthread_join(tid, NULL);
- tid1 = start_thread_on(reader, NULL, dest_cpu);
- pthread_join(tid1, NULL);
- clock_gettime(CLOCK_REALTIME, &ipi_wakeup.time_end);
- ipi_wakeup.latency_ns = (ipi_wakeup.time_end.tv_sec -
ipi_wakeup.time_start.tv_sec) * 1000000000ULL +ipi_wakeup.time_end.tv_nsec - ipi_wakeup.time_start.tv_nsec;- if (baseline)
pthread_cancel(baseline_tid);+}
+static void ipi_test(int src_cpu) +{
- uint64_t avg_arr[cpus], avg_latency;
- int cpu, state;
- FILE *fptr;
- assert(cpuidle_disable_all_states() == 0);
- if (pipe(pipe_fd))
exit(1);- fptr = fopen(log_file, "a");
- fprintf(fptr, "----IPI TEST----\n");
- fprintf(fptr, "----Baseline IPI Latency----\n");
- fprintf(fptr, "%s %10s %18s\n", "SRC_CPU", "DEST_CPU",
"Baseline_latency(ns)");- /* Run the test as dummy once to stablize */
- ipi_test_once(1, src_cpu, cpu_list[0]);
- for (cpu = 0; cpu < cpus; cpu++) {
ipi_test_once(1, src_cpu, cpu_list[cpu]);fprintf(fptr, "%3d %10d %12ld\n", ipi_wakeup.src_cpu,ipi_wakeup.dest_cpu,ipi_wakeup.latency_ns);avg_arr[cpu] = ipi_wakeup.latency_ns;- }
- avg_latency = average(avg_arr, cpus);
- fprintf(fptr, "Baseline average IPI latency(ns): %ld\n\n", avg_latency);
- for (state = 0; state < idle_states; state++) {
fprintf(fptr, "--Enabling state: %d--\n", state);assert(cpuidle_enable_state(state) == 0);fprintf(fptr, "%s %10s %18s\n", "SRC_CPU", "DEST_CPU","IPI_Latency(ns)");for (cpu = 0; cpu < cpus; cpu++) {/* Allow sufficient cycles to go idle */sleep(1);ipi_test_once(0, src_cpu, cpu_list[cpu]);fprintf(fptr, "%3d %10d %18ld\n", ipi_wakeup.src_cpu,ipi_wakeup.dest_cpu,ipi_wakeup.latency_ns);avg_arr[cpu] = ipi_wakeup.latency_ns;}fprintf(fptr, "Expected Latency(ns): %ld\n",us_to_ns(latency_list[state]));avg_latency = average(avg_arr, cpus);fprintf(fptr, "Observed Average IPI latency(ns): %ld\n\n",avg_latency);assert(cpuidle_disable_state(state) == 0);- }
- assert(cpuidle_enable_all_states() == 0);
- fclose(fptr);
+}
+/* Timer based wakeup latencies */ +static int soak_done; +struct timer_data {
- unsigned int src_cpu;
- uint64_t timeout;
- struct timespec time_start;
- struct timespec time_end;
- uint64_t timeout_diff_ns;
+} timer_wakeup;
+void catch_alarm(int sig) +{
- soak_done = 1;
+}
+static void setup_timer(void) +{
- struct itimerval timer_settings = {};
- int err;
- timer_settings.it_value.tv_usec = timer_wakeup.timeout;
- err = setitimer(ITIMER_REAL, &timer_settings, NULL);
- if (err < 0) {
perror("failed to arm interval timer");exit(1);- }
- signal(SIGALRM, catch_alarm);
- while (!soak_done)
sleep(1);+}
+static void *queue_timer(void *timeout) +{
- timer_wakeup.src_cpu = sched_getcpu();
- timer_wakeup.timeout = (uint64_t)timeout;
- setup_timer();
- return NULL;
+}
+static void timeout_test_once(int baseline, uint64_t timeout, int dest_cpu) +{
- pthread_t tid, baseline_tid;
- if (baseline) {
baseline_tid = start_thread_on(util_full_cpu, NULL, dest_cpu);/* Run process for long enough to gain 100% usage */sleep(2);- }
- clock_gettime(CLOCK_REALTIME, &timer_wakeup.time_start);
Here too, should we move the clock_gettime(time_start) in setup_timer() just before calling setittimer() and clockgettime(time_end) in catch_alarm() ?
- tid = start_thread_on(queue_timer, (void *)timeout, dest_cpu);
- pthread_join(tid, NULL);
- clock_gettime(CLOCK_REALTIME, &timer_wakeup.time_end);
- timer_wakeup.timeout_diff_ns = (timer_wakeup.time_end.tv_sec -
timer_wakeup.time_start.tv_sec) * 1000000000ULL +(timer_wakeup.time_end.tv_nsec - timer_wakeup.time_start.tv_nsec);- if (baseline)
pthread_cancel(baseline_tid);+}
+static void timeout_test(unsigned long timeout) +{
- uint64_t avg_arr[cpus], avg_timeout_diff;
- int state, cpu;
- FILE *fptr;
- assert(cpuidle_disable_all_states() == 0);
- fptr = fopen(log_file, "a");
- fprintf(fptr, "----TIMEOUT TEST----\n");
- fprintf(fptr, "----Baseline Timeout Latency diff----\n");
- fprintf(fptr, "%s %10s\n", "SRC_CPU", "Baseline_Latency(ns)");
- /* Run the test as dummy once to stablize */
- timeout_test_once(1, timeout, cpu_list[0]);
- for (cpu = 0; cpu < cpus; cpu++) {
timeout_test_once(1, timeout, cpu_list[cpu]);fprintf(fptr, "%3d %11ld\n", timer_wakeup.src_cpu,timer_wakeup.timeout_diff_ns);avg_arr[cpu] = timer_wakeup.timeout_diff_ns;- }
- avg_timeout_diff = average(avg_arr, cpus);
- fprintf(fptr, "Baseline Average Timeout diff(ns): %ld\n\n",
avg_timeout_diff);- for (state = 0; state < idle_states; state++) {
fprintf(fptr, "--Enabling state: %d--\n", state);assert(cpuidle_enable_state(state) == 0);fprintf(fptr, "%s %11s\n", "SRC_CPU", "Timeout_Latency(ns)");for (cpu = 0; cpu < cpus; cpu++) {/* Allow sufficient cycles to go idle */sleep(1);timeout_test_once(0, timeout,cpu_list[cpu]);fprintf(fptr, "%3d %15ld\n", timer_wakeup.src_cpu,timer_wakeup.timeout_diff_ns);avg_arr[cpu] = timer_wakeup.timeout_diff_ns;}fprintf(fptr, "Expected Residency(ns): %ld\n",us_to_ns(residency_list[state]));avg_timeout_diff = average(avg_arr, cpus);fprintf(fptr, "Observed Average Timeout diff(ns): %ld\n\n",avg_timeout_diff);assert(cpuidle_disable_state(state) == 0);- }
- assert(cpuidle_enable_all_states() == 0);
- fclose(fptr);
+}
+static struct option options[] = {
- {"output", required_argument, 0, 'o'},
- {"mode", required_argument, 0, 'm'},
+};
+static void usage(void) +{
- fprintf(stderr, "Usage: cpuidle <options>\n\n");
- fprintf(stderr, "\t\t--mode=X\t(quick / full / <num_cpus>)\n");
- fprintf(stderr, "\t\t--output=X\tOutput loc (def: cpuidle.log)\n\n");
+}
+void get_n_random_cpus(int *shf_list, int shf_size, int *list, int n) +{
- /* Shuffle the Online CPU list */
- int i;
- int shuffle_index_list[shf_size];
- for (i = 0; i < shf_size; i++)
shuffle_index_list[i] = i;- for (i = 0; i < shf_size; i++) {
int idx = i + rand() % (shf_size - i);int temp = shuffle_index_list[i];shuffle_index_list[i] = shuffle_index_list[idx];shuffle_index_list[idx] = temp;- }
- /* Pick the first n from the shf_list elements */
- for (i = 0; i < n; i++)
list[i] = shf_list[shuffle_index_list[i]];+}
+int main(int argc, char *argv[]) +{
- int total_cpus, online_cpus, option_index = 0;
- int *online_cpu_list;
- signed char c;
- FILE *fptr;
- if (getuid()) {
fprintf(stderr, "cpuidle latency test must run as root\n");exit(1);- }
- total_cpus = sysconf(_SC_NPROCESSORS_ONLN);
- online_cpu_list = malloc(total_cpus * sizeof(int));
- if (!online_cpu_list) {
perror("Malloc failure");exit(1);- }
- online_cpus = get_online_cpus(&online_cpu_list[0], total_cpus);
- if (!online_cpus) {
perror("Unable to get online CPUS");exit(1);- }
- /* Get one CPU for a quick test */
- cpus = 1;
- cpu_list = malloc(1 * sizeof(int));
- srand(time(NULL));
- cpu_list[0] = online_cpu_list[(rand() % online_cpus) + 1];
- while (1) {
c = getopt_long(argc, argv, "", options, &option_index);if (c == -1)break;switch (c) {case 'o':log_file = optarg;break;case 'm':if (!strcmp(optarg, "full")) {cpu_list = realloc(cpu_list,online_cpus * sizeof(int));memcpy(cpu_list, online_cpu_list,online_cpus * sizeof(int));cpus = online_cpus;} else if (strcmp(optarg, "quick")) {int opt_cpus;opt_cpus = atoi(optarg);if (!opt_cpus) {fprintf(stderr, "Error parsing mode\n");usage();exit(1);}if (opt_cpus > online_cpus) {fprintf(stderr, "Number of CPUS > Online CPUs\n");usage();exit(1);}cpu_list = realloc(cpu_list,opt_cpus * sizeof(int));get_n_random_cpus(online_cpu_list, online_cpus,&cpu_list[0], opt_cpus);cpus = opt_cpus;}break;default:usage();exit(1);}- }
- idle_states = get_idle_state_count(online_cpu_list[0]);
- if (idle_states == -1) {
perror("Unable to get idle states");exit(1);- }
- fptr = fopen(log_file, "w+");
- fprintf(fptr, "cpuidle latency selftests. IPI & Timers\n");
- fprintf(fptr, "Number of CPUS: %d\n", total_cpus);
- fprintf(fptr, "Number of idle states: %d\n", idle_states);
- fclose(fptr);
- latency_list = malloc(idle_states * sizeof(uint64_t));
- residency_list = malloc(idle_states * sizeof(uint64_t));
- if (!latency_list || !residency_list) {
perror("Malloc failure");exit(1);- }
- get_latency(online_cpu_list[0]);
- get_residency(online_cpu_list[0]);
- ipi_test(online_cpu_list[0]);
- printf("IPI test done\n");
- fflush(stdout);
- timeout_test(TIMEOUT_US);
- printf("Timeout test done\n");
- fflush(stdout);
- printf("Output logged at: %s\n", log_file);
- free(latency_list);
- free(residency_list);
- free(cpu_list);
- free(online_cpu_list);
- return 0;
+} diff --git a/tools/testing/selftests/cpuidle/settings b/tools/testing/selftests/cpuidle/settings new file mode 100644 index 000000000000..e7b9417537fb --- /dev/null +++ b/tools/testing/selftests/cpuidle/settings @@ -0,0 +1 @@
+timeout=0
2.26.2
On Wed, Sep 02, 2020 at 05:15:05PM +0530, Pratik Rajesh Sampat wrote:
Changelog v3-->v4:
- Overhaul in implementation from kernel module to a userspace selftest
The patch series introduces a mechanism to measure wakeup latency for IPI and timer based interrupts The motivation behind this series is to find significant deviations behind advertised latency and residency values
To achieve this in the userspace, IPI latencies are calculated by sending information through pipes and inducing a wakeup, similarly alarm events are setup for calculate timer based wakeup latencies.
To account for delays from kernel-userspace interactions baseline observations are taken on a 100% busy CPU and subsequent obervations must be considered relative to that.
In theory, wakeups induced by IPI and Timers should have similar wakeup latencies, however in practice there may be deviations which may need to be captured.
One downside of the userspace approach in contrast to the kernel implementation is that the run to run variance can turn out to be high in the order of ms; which is the scope of the experiments at times.
Another downside of the userspace approach is that it takes much longer to run and hence a command-line option quick and full are added to make sure quick 1 CPU tests can be carried out when needed and otherwise it can carry out a full system comprehensive test.
Usage
./cpuidle --mode <full / quick / num_cpus> --output <output location> full: runs on all CPUS quick: run on a random CPU num_cpus: Limit the number of CPUS to run on
Sample output snippet
--IPI Latency Test--- SRC_CPU DEST_CPU IPI_Latency(ns) ... 0 5 256178 0 6 478161 0 7 285445 0 8 273553 Expected IPI latency(ns): 100000 Observed Average IPI latency(ns): 248334
I suppose by run-to-run variance you are referring to the outliers in the above sequence (like 478161) ? Or is it that each time you run your test program you observe completely different series of values ?
If it is the former, then perhaps we could discard the outliers for the purpose of average latency computation and print the max, min and the corrected-average values above.
--Timeout Latency Test-- --Baseline Timeout Latency measurement: CPU Busy-- Wakeup_src Baseline_delay(ns) ... 32 972405 33 1004287 34 986663 35 994022 Expected timeout(ns): 10000000 Observed Average timeout diff(ns): 991844
It would be good to see a complete sample output, perhaps for the --mode=10 so that it is easy to discern if there are cases when the observed timeouts/IPI latencies for the busy case are larger than the idle-case.
Pratik Rajesh Sampat (1): selftests/cpuidle: Add support for cpuidle latency measurement
tools/testing/selftests/Makefile | 1 + tools/testing/selftests/cpuidle/Makefile | 7 + tools/testing/selftests/cpuidle/cpuidle.c | 616 ++++++++++++++++++++++ tools/testing/selftests/cpuidle/settings | 1 + 4 files changed, 625 insertions(+) create mode 100644 tools/testing/selftests/cpuidle/Makefile create mode 100644 tools/testing/selftests/cpuidle/cpuidle.c create mode 100644 tools/testing/selftests/cpuidle/settings
-- 2.26.2
-- Thanks and Regards gautham.
linux-kselftest-mirror@lists.linaro.org
-
Artem Bityutskiy -
Gautham R Shenoy -
Pratik Rajesh Sampat -
Pratik Sampat