Use plain Makefile instead of CMake

Move each test to a single binary.

.gitignore

@@ -1 +1,2 @@
-build
+data
+bench6.*

CMakeLists.txt

@@ -1,30 +0,0 @@
-# Copyright (c) 2022 Barcelona Supercomputing Center (BSC)
-# SPDX-License-Identifier: GPL-3.0-or-later
-
-cmake_minimum_required(VERSION 3.10)
-
-project(BENCH6 LANGUAGES C)
-
-add_compile_options(-Wall -Wextra -Wformat
-  -Wmissing-prototypes -Wstrict-prototypes
-  #-Wconversion -Wsign-conversion
-  -Wold-style-definition -pedantic
-  -Werror
-)
-
-set(CMAKE_C_COMPILER "clang")
-
-set(CMAKE_C_STANDARD 11)
-set(CMAKE_C_EXTENSIONS FALSE)
-
-add_executable(bench6
-  src/main.c
-  src/common.c
-  src/sched_get.c
-  src/sched_add.c
-  src/register_deps.c)
-
-target_compile_options(bench6 PRIVATE -fompss-2)
-target_link_options(bench6 PRIVATE -fompss-2)
-
-install(TARGETS bench6 RUNTIME DESTINATION bin)

Makefile

@@ -0,0 +1,23 @@
+CC=clang
+CFLAGS=-O3 -fompss-2
+
+BENCHMARKS=\
+	sched_get \
+	sched_add \
+	register_deps
+
+BIN=$(addprefix bench6.,$(BENCHMARKS)) 
+DATA=$(addsuffix .csv, $(addprefix data/,$(BENCHMARKS)))
+PLOT=$(DATA:=.png)
+
+all: $(BIN) $(DATA) $(PLOT)
+
+bench6.%: src/%.c src/common.c
+	$(CC) $(CFLAGS) -o $@ $^
+
+data/%.csv.png: data/%.csv plot/%.R
+	Rscript plot/$(*F).R $<
+
+data/%.csv: bench6.%
+	mkdir -p data
+	./$^ > $@

plot/register_deps.R

@@ -0,0 +1,34 @@
+library(ggplot2)
+library(dplyr, warn.conflicts = FALSE)
+library(scales)
+library(jsonlite)
+library(readr)
+
+# Load the arguments (argv)
+args = commandArgs(trailingOnly=TRUE)
+
+input_file = args[1]
+
+df = read_delim(input_file, delim=",", comment="#") %>%
+  mutate(ndeps = as.factor(ndeps)) %>%
+  mutate(time_per_task = time_per_task * 1e6) %>%
+  group_by(ndeps) %>%
+  mutate(median_time = median(time_per_task)) %>%
+  ungroup()
+
+dpi = 300
+h = 6
+w = 15
+
+# ---------------------------------------------------------------------
+
+p = ggplot(df, aes(x=ndeps, y=time_per_task)) +
+  geom_boxplot() +
+  theme_bw() +
+  labs(
+    x="Number of dependencies per task",
+    y="Creation and registration time per task (us)",
+    title="bench6.register_deps: registration time vs number of dependencies")
+
+ggsave(sprintf("%s.png", input_file), plot=p, width=w, height=h, dpi=dpi)
+

plot/sched_add.R

@@ -0,0 +1,39 @@
+library(ggplot2)
+library(dplyr, warn.conflicts = FALSE)
+library(scales)
+library(jsonlite)
+library(readr)
+
+# Load the arguments (argv)
+args = commandArgs(trailingOnly=TRUE)
+
+input_file = args[1]
+
+df = read_delim(input_file, delim=",", comment="#") %>%
+  mutate(run = as.factor(run)) %>%
+  mutate(time_per_task_per_cpu = time_per_task_per_cpu * 1e9)
+
+dpi = 150
+h = 2
+w = 6
+
+# ---------------------------------------------------------------------
+
+#p = ggplot(df, aes(x=run, y=time_per_task)) +
+#  geom_point() +
+#  theme_bw() +
+#  labs(
+#       x = "Number of run",
+#       y="get_ready_task() time (ns)",
+#       title="Nanos6: get ready task time")
+
+p = ggplot(df, aes(x=time_per_task_per_cpu)) +
+  geom_boxplot() +
+  theme_bw() +
+  scale_x_continuous(breaks = breaks_pretty(10)) +
+  labs(
+       x="Duration per task per CPU (ns / task * CPU)",
+       title="bench6.sched_add: time to unblock N tasks")
+
+ggsave(sprintf("%s.png", input_file), plot=p, width=w, height=h, dpi=dpi)
+

plot/sched_get.R

@@ -0,0 +1,39 @@
+library(ggplot2)
+library(dplyr, warn.conflicts = FALSE)
+library(scales)
+library(jsonlite)
+library(readr)
+
+# Load the arguments (argv)
+args = commandArgs(trailingOnly=TRUE)
+
+input_file = args[1]
+
+df = read_delim(input_file, delim=",", comment="#") %>%
+  mutate(run = as.factor(run)) %>%
+  mutate(time_per_task_per_cpu = time_per_task_per_cpu * 1e9)
+
+dpi = 150
+h = 2
+w = 6
+
+# ---------------------------------------------------------------------
+
+#p = ggplot(df, aes(x=run, y=time_per_task_per_cpu)) +
+#  geom_point() +
+#  theme_bw() +
+#  labs(
+#       x = "Number of run",
+#       y="get_ready_task() time (ns)",
+#       title="Nanos6: get ready task time")
+
+p = ggplot(df, aes(x=time_per_task_per_cpu)) +
+  geom_boxplot() +
+  theme_bw() +
+  scale_x_continuous(breaks = breaks_pretty(10)) +
+  labs(
+       x="Average duration per task per CPU (ns / task * CPU)",
+       title="bench6.sched_get: average time to schedule N ready tasks")
+
+ggsave(sprintf("%s.png", input_file), plot=p, width=w, height=h, dpi=dpi)
+

src/bench6.h

@@ -7,6 +7,7 @@
 #define UNUSED(x) (void)(x)
 
 double get_time(void);
+int get_ncpus();
 
 int bench6_creator(int argc, char *argv[]);
 int bench6_sched_get(int argc, char *argv[]);

src/common.c

@@ -5,6 +5,7 @@
 
 #include "bench6.h"
 
+#include <nanos6/debug.h>
 #include <time.h>
 #include <stdio.h>
 #include <stdlib.h>
@@ -22,3 +23,8 @@ double get_time()
 	return (double)(tv.tv_sec) +
 		(double)tv.tv_nsec * 1.0e-9;
 }
+
+int get_ncpus()
+{
+	return (int) nanos6_get_num_cpus();
+}

src/register_deps.c

@@ -18,7 +18,8 @@ static long ntasks = 100;
 static int
 usage(char *argv[])
 {
-	fprintf(stderr, "Usage: bench6 %s [-r NRUNS] [-t NTASKS]\n", argv[0]);
+	fprintf(stderr, "Bench6: A set of Nanos6 micro-benchmarks\n");
+	fprintf(stderr, "Usage: %s [-r NRUNS] [-t NTASKS]\n", argv[0]);
 	fprintf(stderr, "\n");
 	fprintf(stderr,
 "Measure the time it takes to create and register NTASKS tasks\n"
@@ -49,7 +50,7 @@ do_run(int run)
 }
 
 int
-bench6_register_deps(int argc, char *argv[])
+main(int argc, char *argv[])
 {
 	int opt;
 

src/sched_add.c

@@ -13,6 +13,7 @@
 #include <time.h>
 #include <unistd.h>
 
+static int ncpus = -1;
 static long nruns = 30L;
 static long ntasks = 10000L;
 
@@ -22,7 +23,8 @@ static void **handle;
 static int
 usage(char *argv[])
 {
-	fprintf(stderr, "Usage: bench6 %s [-r NRUNS] [-t NTASKS]\n", argv[0]);
+	fprintf(stderr, "Bench6: A set of Nanos6 micro-benchmarks\n");
+	fprintf(stderr, "Usage: %s [-r NRUNS] [-t NTASKS]\n", argv[0]);
 	fprintf(stderr, "\n");
 	fprintf(stderr,
 "Measure the time it takes to unblock NTASKS which will end\n"
@@ -66,11 +68,11 @@ do_run(int run)
 	double t1 = get_time();
 	printf("%d,%ld,%e,%e\n",
 			run, ntasks, (t1 - t0),
-			(t1 - t0) / ((double) ntasks));
+			(t1 - t0) / ((double) ntasks) * ((double) ncpus));
 }
 
 int
-bench6_sched_add(int argc, char *argv[])
+main(int argc, char *argv[])
 {
 	int opt;
 	while ((opt = getopt(argc, argv, "hr:t:")) != -1) {
@@ -87,6 +89,8 @@ bench6_sched_add(int argc, char *argv[])
 		}
 	}
 
+	ncpus = get_ncpus();
+
 	handle = calloc(ntasks, sizeof(void *));
 
 	if (handle == NULL) {
@@ -94,7 +98,7 @@ bench6_sched_add(int argc, char *argv[])
 		return -1;
 	}
 
-	printf("%s,%s,%s,%s\n", "run", "ntasks", "time_us", "time_per_task");
+	printf("%s,%s,%s,%s\n", "run", "ntasks", "time", "time_per_task_per_cpu");
 	for (int run = 0; run < nruns; run++)
 		do_run(run);
 

src/sched_get.c

@@ -10,6 +10,7 @@
 #include <stdlib.h>
 #include <unistd.h>
 
+static int ncpus = -1;
 static long nruns = 30L;
 static long ntasks = 20000L;
 
@@ -18,7 +19,8 @@ static atomic_int wait = 0;
 static int
 usage(char *argv[])
 {
-	fprintf(stderr, "Usage: bench6 %s [-r NRUNS] [-t NTASKS]\n", argv[0]);
+	fprintf(stderr, "Bench6: A set of Nanos6 micro-benchmarks\n");
+	fprintf(stderr, "Usage: %s [-r NRUNS] [-t NTASKS]\n", argv[0]);
 	fprintf(stderr, "\n");
 	fprintf(stderr,
 "Creates NTASKS tasks without dependencies, but the tasks don't\n"
@@ -48,13 +50,13 @@ do_run(int run)
 
 	#pragma oss taskwait
 	double t1 = get_time();
-	printf("%d,%ld,%e,%e\n",
-			run, ntasks, (t1 - t0),
-			(t1 - t0) / ((double) ntasks));
+	printf("%d,%ld,%d,%e,%e\n",
+			run, ntasks, ncpus, (t1 - t0),
+			(t1 - t0) / ((double) ntasks) * ((double) ncpus));
 }
 
 int
-bench6_sched_get(int argc, char *argv[])
+main(int argc, char *argv[])
 {
 	int opt;
 
@@ -72,7 +74,9 @@ bench6_sched_get(int argc, char *argv[])
 		}
 	}
 
-	printf("%s,%s,%s,%s\n", "run", "ntasks", "time", "time_per_task");
+	ncpus = get_ncpus();
+
+	printf("%s,%s,%s,%s,%s\n", "run", "ntasks", "ncpus", "time", "time_per_task_per_cpu");
 	for (int run = 0; run < nruns; run++)
 		do_run(run);