/* ** Copyright 2010 The Android Open Source Project ** ** Licensed under the Apache License, Version 2.0 (the "License"); ** you may not use this file except in compliance with the License. ** You may obtain a copy of the License at ** ** http://www.apache.org/licenses/LICENSE-2.0 ** ** Unless required by applicable law or agreed to in writing, software ** distributed under the License is distributed on an "AS IS" BASIS, ** WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. ** See the License for the specific language governing permissions and ** limitations under the License. */ /* * Micro-benchmarking of sleep/cpu speed/memcpy/memset/memory reads. */ #include <stdio.h> #include <stdlib.h> #include <ctype.h> #include <math.h> #include <sched.h> #include <sys/resource.h> #include <time.h> #include <unistd.h> // The default size of data that will be manipulated in each iteration of // a memory benchmark. Can be modified with the --data_size option. #define DEFAULT_DATA_SIZE 1000000000 // Number of nanoseconds in a second. #define NS_PER_SEC 1000000000 // The maximum number of arguments that a benchmark will accept. #define MAX_ARGS 2 // Use macros to compute values to try and avoid disturbing memory as much // as possible after each iteration. #define COMPUTE_AVERAGE_KB(avg_kb, bytes, time_ns) \ avg_kb = ((bytes) / 1024.0) / ((double)(time_ns) / NS_PER_SEC); #define COMPUTE_RUNNING(avg, running_avg, square_avg, cur_idx) \ running_avg = ((running_avg) / ((cur_idx) + 1)) * (cur_idx) + (avg) / ((cur_idx) + 1); \ square_avg = ((square_avg) / ((cur_idx) + 1)) * (cur_idx) + ((avg) / ((cur_idx) + 1)) * (avg); #define GET_STD_DEV(running_avg, square_avg) \ sqrt((square_avg) - (running_avg) * (running_avg)) // Contains information about benchmark options. typedef struct { bool print_average; bool print_each_iter; int dst_align; int src_align; int cpu_to_lock; int data_size; int args[MAX_ARGS]; int num_args; } command_data_t; // Struct that contains a mapping of benchmark name to benchmark function. typedef struct { const char *name; int (*ptr)(const command_data_t &cmd_data); } function_t; // Get the current time in nanoseconds. uint64_t nanoTime() { struct timespec t; t.tv_sec = t.tv_nsec = 0; clock_gettime(CLOCK_MONOTONIC, &t); return static_cast<uint64_t>(t.tv_sec) * NS_PER_SEC + t.tv_nsec; } // Allocate memory with a specific alignment and return that pointer. // This function assumes an alignment value that is a power of 2. // If the alignment is 0, then use the pointer returned by malloc. uint8_t *allocateAlignedMemory(size_t size, int alignment) { uint64_t ptr = reinterpret_cast<uint64_t>(malloc(size + 2 * alignment)); if (!ptr) return NULL; if (alignment > 0) { // When setting the alignment, set it to exactly the alignment chosen. // The pointer returned will be guaranteed not to be aligned to anything // more than that. ptr += alignment - (ptr & (alignment - 1)); ptr |= alignment; } return reinterpret_cast<uint8_t*>(ptr); } int benchmarkSleep(const command_data_t &cmd_data) { uint64_t time_ns; int delay = cmd_data.args[0]; int iters = cmd_data.args[1]; bool print_each_iter = cmd_data.print_each_iter; bool print_average = cmd_data.print_average; double avg, running_avg = 0.0, square_avg = 0.0; for (int i = 0; iters == -1 || i < iters; i++) { time_ns = nanoTime(); sleep(delay); time_ns = nanoTime() - time_ns; avg = (double)time_ns / NS_PER_SEC; if (print_average) { COMPUTE_RUNNING(avg, running_avg, square_avg, i); } if (print_each_iter) { printf("sleep(%d) took %.06f seconds\n", delay, avg); } } if (print_average) { printf(" sleep(%d) average %.06f seconds std dev %f\n", delay, running_avg, GET_STD_DEV(running_avg, square_avg)); } return 0; } int benchmarkCpu(const command_data_t &cmd_data) { // Use volatile so that the loop is not optimized away by the compiler. volatile int cpu_foo; uint64_t time_ns; int iters = cmd_data.args[1]; bool print_each_iter = cmd_data.print_each_iter; bool print_average = cmd_data.print_average; double avg, running_avg = 0.0, square_avg = 0.0; for (int i = 0; iters == -1 || i < iters; i++) { time_ns = nanoTime(); for (cpu_foo = 0; cpu_foo < 100000000; cpu_foo++); time_ns = nanoTime() - time_ns; avg = (double)time_ns / NS_PER_SEC; if (print_average) { COMPUTE_RUNNING(avg, running_avg, square_avg, i); } if (print_each_iter) { printf("cpu took %.06f seconds\n", avg); } } if (print_average) { printf(" cpu average %.06f seconds std dev %f\n", running_avg, GET_STD_DEV(running_avg, square_avg)); } return 0; } int benchmarkMemset(const command_data_t &cmd_data) { int size = cmd_data.args[0]; int iters = cmd_data.args[1]; uint8_t *dst = allocateAlignedMemory(size, cmd_data.dst_align); if (!dst) return -1; double avg_kb, running_avg_kb = 0.0, square_avg_kb = 0.0; uint64_t time_ns; int j; bool print_average = cmd_data.print_average; bool print_each_iter = cmd_data.print_each_iter; int copies = cmd_data.data_size/size; for (int i = 0; iters == -1 || i < iters; i++) { time_ns = nanoTime(); for (j = 0; j < copies; j++) memset(dst, 0, size); time_ns = nanoTime() - time_ns; // Compute in kb to avoid any overflows. COMPUTE_AVERAGE_KB(avg_kb, copies * size, time_ns); if (print_average) { COMPUTE_RUNNING(avg_kb, running_avg_kb, square_avg_kb, i); } if (print_each_iter) { printf("memset %dx%d bytes took %.06f seconds (%f MB/s)\n", copies, size, (double)time_ns / NS_PER_SEC, avg_kb / 1024.0); } } if (print_average) { printf(" memset %dx%d bytes average %.2f MB/s std dev %.4f\n", copies, size, running_avg_kb / 1024.0, GET_STD_DEV(running_avg_kb, square_avg_kb) / 1024.0); } return 0; } int benchmarkMemcpy(const command_data_t &cmd_data) { int size = cmd_data.args[0]; int iters = cmd_data.args[1]; uint8_t *src = allocateAlignedMemory(size, cmd_data.src_align); if (!src) return -1; uint8_t *dst = allocateAlignedMemory(size, cmd_data.dst_align); if (!dst) return -1; uint64_t time_ns; double avg_kb, running_avg_kb = 0.0, square_avg_kb = 0.0; int j; bool print_average = cmd_data.print_average; bool print_each_iter = cmd_data.print_each_iter; int copies = cmd_data.data_size / size; for (int i = 0; iters == -1 || i < iters; i++) { time_ns = nanoTime(); for (j = 0; j < copies; j++) memcpy(dst, src, size); time_ns = nanoTime() - time_ns; // Compute in kb to avoid any overflows. COMPUTE_AVERAGE_KB(avg_kb, copies * size, time_ns); if (print_average) { COMPUTE_RUNNING(avg_kb, running_avg_kb, square_avg_kb, i); } if (print_each_iter) { printf("memcpy %dx%d bytes took %.06f seconds (%f MB/s)\n", copies, size, (double)time_ns / NS_PER_SEC, avg_kb / 1024.0); } } if (print_average) { printf(" memcpy %dx%d bytes average %.2f MB/s std dev %.4f\n", copies, size, running_avg_kb/1024.0, GET_STD_DEV(running_avg_kb, square_avg_kb) / 1024.0); } return 0; } int benchmarkMemread(const command_data_t &cmd_data) { int size = cmd_data.args[0]; int iters = cmd_data.args[1]; int *src = reinterpret_cast<int*>(malloc(size)); if (!src) return -1; // Use volatile so the compiler does not optimize away the reads. volatile int foo; uint64_t time_ns; int j, k; double avg_kb, running_avg_kb = 0.0, square_avg_kb = 0.0; bool print_average = cmd_data.print_average; bool print_each_iter = cmd_data.print_each_iter; int c = cmd_data.data_size / size; for (int i = 0; iters == -1 || i < iters; i++) { time_ns = nanoTime(); for (j = 0; j < c; j++) for (k = 0; k < size/4; k++) foo = src[k]; time_ns = nanoTime() - time_ns; // Compute in kb to avoid any overflows. COMPUTE_AVERAGE_KB(avg_kb, c * size, time_ns); if (print_average) { COMPUTE_RUNNING(avg_kb, running_avg_kb, square_avg_kb, i); } if (print_each_iter) { printf("read %dx%d bytes took %.06f seconds (%f MB/s)\n", c, size, (double)time_ns / NS_PER_SEC, avg_kb / 1024.0); } } if (print_average) { printf(" read %dx%d bytes average %.2f MB/s std dev %.4f\n", c, size, running_avg_kb/1024.0, GET_STD_DEV(running_avg_kb, square_avg_kb) / 1024.0); } return 0; } // Create the mapping structure. function_t function_table[] = { { "sleep", benchmarkSleep }, { "cpu", benchmarkCpu }, { "memset", benchmarkMemset }, { "memcpy", benchmarkMemcpy }, { "memread", benchmarkMemread }, { NULL, NULL } }; void usage() { printf("Usage:\n"); printf(" micro_bench [--data_size DATA_BYTES] [--print_average]\n"); printf(" [--no_print_each_iter] [--lock_to_cpu CORE]\n"); printf(" --data_size DATA_BYTES\n"); printf(" For the data benchmarks (memcpy/memset/memread) the approximate\n"); printf(" size of data, in bytes, that will be manipulated in each iteration.\n"); printf(" --print_average\n"); printf(" Print the average and standard deviation of all iterations.\n"); printf(" --no_print_each_iter\n"); printf(" Do not print any values in each iteration.\n"); printf(" --lock_to_cpu CORE\n"); printf(" Lock to the specified CORE. The default is to use the last core found.\n"); printf(" ITERS\n"); printf(" The number of iterations to execute each benchmark. If not\n"); printf(" passed in then run forever.\n"); printf(" micro_bench sleep TIME_TO_SLEEP [ITERS]\n"); printf(" TIME_TO_SLEEP\n"); printf(" The time in seconds to sleep.\n"); printf(" micro_bench cpu UNUSED [ITERS]\n"); printf(" micro_bench [--dst_align ALIGN] memset NUM_BYTES [ITERS]\n"); printf(" --dst_align ALIGN\n"); printf(" Align the memset destination pointer to ALIGN. The default is to use the\n"); printf(" value returned by malloc.\n"); printf(" micro_bench [--src_align ALIGN] [--dst_align ALIGN] memcpy NUM_BYTES [ITERS]\n"); printf(" --src_align ALIGN\n"); printf(" Align the memcpy source pointer to ALIGN. The default is to use the\n"); printf(" value returned by malloc.\n"); printf(" --dst_align ALIGN\n"); printf(" Align the memcpy destination pointer to ALIGN. The default is to use the\n"); printf(" value returned by malloc.\n"); printf(" micro_bench memread NUM_BYTES [ITERS]\n"); } function_t *processOptions(int argc, char **argv, command_data_t *cmd_data) { function_t *command = NULL; // Initialize the command_flags. cmd_data->print_average = false; cmd_data->print_each_iter = true; cmd_data->dst_align = 0; cmd_data->src_align = 0; cmd_data->num_args = 0; cmd_data->cpu_to_lock = -1; cmd_data->data_size = DEFAULT_DATA_SIZE; for (int i = 0; i < MAX_ARGS; i++) { cmd_data->args[i] = -1; } for (int i = 1; i < argc; i++) { if (argv[i][0] == '-') { int *save_value = NULL; if (strcmp(argv[i], "--print_average") == 0) { cmd_data->print_average = true; } else if (strcmp(argv[i], "--no_print_each_iter") == 0) { cmd_data->print_each_iter = false; } else if (strcmp(argv[i], "--dst_align") == 0) { save_value = &cmd_data->dst_align; } else if (strcmp(argv[i], "--src_align") == 0) { save_value = &cmd_data->src_align; } else if (strcmp(argv[i], "--lock_to_cpu") == 0) { save_value = &cmd_data->cpu_to_lock; } else if (strcmp(argv[i], "--data_size") == 0) { save_value = &cmd_data->data_size; } else { printf("Unknown option %s\n", argv[i]); return NULL; } if (save_value) { // Checking both characters without a strlen() call should be // safe since as long as the argument exists, one character will // be present (\0). And if the first character is '-', then // there will always be a second character (\0 again). if (i == argc - 1 || (argv[i + 1][0] == '-' && !isdigit(argv[i + 1][1]))) { printf("The option %s requires one argument.\n", argv[i]); return NULL; } *save_value = atoi(argv[++i]); } } else if (!command) { for (function_t *function = function_table; function->name != NULL; function++) { if (strcmp(argv[i], function->name) == 0) { command = function; break; } } if (!command) { printf("Uknown command %s\n", argv[i]); return NULL; } } else if (cmd_data->num_args > MAX_ARGS) { printf("More than %d number arguments passed in.\n", MAX_ARGS); return NULL; } else { cmd_data->args[cmd_data->num_args++] = atoi(argv[i]); } } // Check the arguments passed in make sense. if (cmd_data->num_args != 1 && cmd_data->num_args != 2) { printf("Not enough arguments passed in.\n"); return NULL; } else if (cmd_data->dst_align < 0) { printf("The --dst_align option must be greater than or equal to 0.\n"); return NULL; } else if (cmd_data->src_align < 0) { printf("The --src_align option must be greater than or equal to 0.\n"); return NULL; } else if (cmd_data->data_size <= 0) { printf("The --data_size option must be a positive number.\n"); return NULL; } else if ((cmd_data->dst_align & (cmd_data->dst_align - 1))) { printf("The --dst_align option must be a power of 2.\n"); return NULL; } else if ((cmd_data->src_align & (cmd_data->src_align - 1))) { printf("The --src_align option must be a power of 2.\n"); return NULL; } return command; } bool raisePriorityAndLock(int cpu_to_lock) { cpu_set_t cpuset; if (setpriority(PRIO_PROCESS, 0, -20)) { perror("Unable to raise priority of process.\n"); return false; } CPU_ZERO(&cpuset); if (sched_getaffinity(0, sizeof(cpuset), &cpuset) != 0) { perror("sched_getaffinity failed"); return false; } if (cpu_to_lock < 0) { // Lock to the last active core we find. for (int i = 0; i < CPU_SETSIZE; i++) { if (CPU_ISSET(i, &cpuset)) { cpu_to_lock = i; } } } else if (!CPU_ISSET(cpu_to_lock, &cpuset)) { printf("Cpu %d does not exist.\n", cpu_to_lock); return false; } if (cpu_to_lock < 0) { printf("Cannot find any valid cpu to lock.\n"); return false; } CPU_ZERO(&cpuset); CPU_SET(cpu_to_lock, &cpuset); if (sched_setaffinity(0, sizeof(cpuset), &cpuset) != 0) { perror("sched_setaffinity failed"); return false; } return true; } int main(int argc, char **argv) { command_data_t cmd_data; function_t *command = processOptions(argc, argv, &cmd_data); if (!command) { usage(); return -1; } if (!raisePriorityAndLock(cmd_data.cpu_to_lock)) { return -1; } printf("%s\n", command->name); return (*command->ptr)(cmd_data); }