Merge branch 'work-work-work' into 'main'

Work work work

See merge request i4/manycore/fgbs-syscall-eval!3

Makefile

 BENCH_MAIN := bench.c
-SYSCALLS := blocking io-uring io-uring-sqpoll io-uring-no-syscall epoll aio-sig aio-thrd
+SYSCALLS := blocking io-uring io-uring-sqpoll io-uring-no-syscall epoll paio-sig paio-thrd
 
 OBJ := $(addprefix bench-,$(SYSCALLS))
 
 LDFLAGS := -luring -pthread -lrt
 
-CFLAGS := -Werror -Wall -g -O3
+CFLAGS := -Werror -Wall -g -O3 -D_GNU_SOURCE
 # CFLAGS := -Werror -Wall -g -O0
 
 .PHONY: all clean eval docker-eval check

bench.c

 #include <err.h>
 #include <fcntl.h>
+#include <sched.h>
 #include <stdint.h>
 #include <stdio.h>
 #include <stdlib.h>
 #include <sys/eventfd.h>
+#include <unistd.h>
 
 #include "stopwatch.h"
 
 void init(int fd);
 void do_write(int fd, const void *buf, size_t count);
 
-const size_t warmup = 10000;
-const size_t iterations = 1000000;
+size_t warmup = 10000;
+size_t iterations = 1000000;
 
 #define BUFSIZE 64
 
@@ -23,18 +25,33 @@ static int create_eventfd() {
 	return fd;
 }
 
+static void set_cpu_affinity(int cpu) {
+	cpu_set_t set;
+	// NOLINTNEXTLINE(clang-analyzer-security.insecureAPI.DeprecatedOrUnsafeBufferHandling)
+	CPU_ZERO(&set);
+	CPU_SET(cpu, &set);
+	if (sched_setaffinity(getpid(), sizeof(set), &set) == -1)
+		err(EXIT_FAILURE, "sched_setaffinity failed");
+}
+
 int main() {
 	uint64_t cycles_sum = 0;
 	uint64_t nanos_sum = 0;
 	uint64_t write_buf = 1;
 
+	set_cpu_affinity(0);
+
 	int fd = create_eventfd();
 
 	init(fd);
+	const size_t exp_warmup = warmup;
+	const size_t exp_iterations = iterations;
+
+	if (exp_iterations == 0) errx(EXIT_FAILURE, "experiment must do at least one iteration");
 
-	for (size_t i = 0; i < warmup; ++i) do_write(fd, &write_buf, sizeof(write_buf));
+	for (size_t i = 0; i < exp_warmup; ++i) do_write(fd, &write_buf, sizeof(write_buf));
 
-	for (int64_t i = 1; i <= iterations; ++i) {
+	for (int64_t i = 1; i <= exp_iterations; ++i) {
 		do_write(fd, &write_buf, sizeof(write_buf));
 		if (__builtin_add_overflow(nanos_sum, clock_diff_nanos(), &nanos_sum))
 			errx(EXIT_FAILURE, "nanos overflowed at %ld", i);
@@ -45,8 +62,8 @@ int main() {
 	// Since uint64_t <-> double conversion are not well defined
 	// and we use really small units (cycles and nanoseconds) I am willing
 	// to accept that we throw away anything after the decimal point.
-	uint64_t avg_nanos = nanos_sum / iterations;
-	uint64_t avg_cycles = cycles_sum / iterations;
+	uint64_t avg_nanos = nanos_sum / exp_iterations;
+	uint64_t avg_cycles = cycles_sum / exp_iterations;
 
 	printf("%lu ns, %lu cycles\n", avg_nanos, avg_cycles);
 	return 0;

common.h

 #pragma once
 
 #define unlikely(x) __builtin_expect(!!(x), 0)
+
+extern size_t warmup, iterations;

aio-thrd.c → paio-thrd.c

@@ -5,6 +5,7 @@
 #include <stdint.h>
 #include <stdlib.h>
 
+#include "common.h"
 #include "stopwatch.h"
 
 atomic_int done;
@@ -19,6 +20,9 @@ void callback(union sigval sigval) {
 }
 
 void init(int fd) {
+	// decrease the experiment size because this takes much longer than anything else
+	warmup = 10;
+	iterations = 100;
 	aiocb.aio_fildes = fd;
 	aiocb.aio_sigevent.sigev_notify = SIGEV_THREAD;
 	aiocb.aio_sigevent.sigev_notify_function = callback;