Simoop: try to simulate the MM and FS loads we see from hadoop

Signed-off-by: Chris Mason <clm@fb.com>

2 files changed, 1343 insertions, 0 deletions

diff --git a/Makefile b/Makefile
new file mode 100644
index 0000000..b2221f3
--- /dev/null
+++ b/Makefile
@@ -0,0 +1,27 @@
+CC      = gcc
+CFLAGS  = -Wall -O2 -g -W
+ALL_CFLAGS = $(CFLAGS) -D_GNU_SOURCE -D_LARGEFILE_SOURCE -D_FILE_OFFSET_BITS=64
+
+PROGS = simoop
+ALL = $(PROGS)
+
+$(PROGS): | depend
+
+all: $(ALL)
+
+%.o: %.c
+	$(CC) -o $*.o -c $(ALL_CFLAGS) $<
+
+simoop: simoop.o
+	$(CC) $(ALL_CFLAGS) -o $@ $(filter %.o,$^) -lpthread
+
+depend:
+	@$(CC) -MM $(ALL_CFLAGS) *.c 1> .depend
+
+clean:
+	-rm -f *.o $(PROGS) .depend
+
+ifneq ($(wildcard .depend),)
+include .depend
+endif
+
diff --git a/simoop.c b/simoop.c
new file mode 100644
index 0000000..2069c08
--- /dev/null
+++ b/simoop.c
@@ -0,0 +1,1316 @@
+/*
+ * simoop.c
+ *
+ * Copyright (C) 2016 Facebook
+ * Chris Mason <clm@fb.com>
+ *
+ * GPLv2, portions copied from the kernel and from Jens Axboe's fio
+ */
+#include <stdio.h>
+#include <stdlib.h>
+#include <pthread.h>
+#include <fcntl.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <unistd.h>
+#include <errno.h>
+#include <getopt.h>
+#include <sys/time.h>
+#include <time.h>
+#include <string.h>
+#include <linux/futex.h>
+#include <sys/syscall.h>
+#include <sys/mman.h>
+#include <libgen.h>
+#include <locale.h>
+
+/* these are part of the histogram accounting */
+#define PLAT_BITS	8
+#define PLAT_VAL	(1 << PLAT_BITS)
+#define PLAT_GROUP_NR	19
+#define PLAT_NR		(PLAT_GROUP_NR * PLAT_VAL)
+#define PLAT_LIST_MAX	20
+
+/* how deep a directory chain to make */
+#define DIR_LEVEL 64
+
+/* buffer size for reads and writes */
+#define BUF_SIZE 65536
+
+/*
+ * we make a few different kinds of files, these are appended onto the
+ * file name to separate them
+ */
+#define DATA_FILE NULL
+#define RESULT_FILE "extra"
+#define TMP_FILE "tmp"
+
+/* each path in the paths array gets a thread pool hammering on it. */
+char **paths;
+int total_paths = 0;
+
+/* -t  number of workers thread */
+static int worker_threads = 16;
+/* -r  seconds */
+static int runtime = 30;
+/* -c  usec */
+static unsigned long long cputime = 3000000;
+/* -f size of the files we create */
+static unsigned long long file_size = 64 * 1024 * 1024;
+/* -n number of files we create */
+static unsigned long num_files = 65536;
+/* -R read size */
+static unsigned long read_size = 2 * 1024 * 1024;
+/* -W write size */
+static unsigned long write_size = 2 * 1024 * 1024;
+/* -T number of files to read */
+static int rw_threads = 8;
+/* -D number of threads running du */
+static int du_threads = 1;
+/* memory to allocate and use during each task */
+static int thinking_mem = 128 * 1024 * 1024;
+/* should we do a truncate and fsync after every write */
+static int funksync = 0;
+
+/* -M how much memory we allocate to benchmark allocations */
+static int mmap_size = 64 * 1024 * 1024;
+
+/* these do nothing but spin */
+static int cpu_threads = 24;
+
+/* how long we sleep while processing requests */
+static int sleeptime = 10000;
+
+/*
+ * after warmup_seconds, we reset the counters to get rid of noise from
+ * early in the run
+ */
+static int warmup_seconds = 60;
+
+/* reporting interval */
+static int interval_seconds = 120;
+
+/* the master thread flips this to true when runtime is up */
+static volatile unsigned long stopping = 0;
+static volatile unsigned long warmup_done = 0;
+
+/*
+ * one stat struct per thread data, when the workers sleep this records the
+ * latency between when they are woken up and when they actually get the
+ * CPU again.  The message threads sum up the stats of all the workers and
+ * then bubble them up to main() for printing
+ */
+struct stats {
+	unsigned int plat[PLAT_NR];
+	unsigned int nr_samples;
+	unsigned int max;
+	unsigned int min;
+};
+
+/* this defines which latency profiles get printed */
+#define PLIST_P99 2
+#define PLIST_P95 1
+#define PLIST_P50 0
+static double plist[PLAT_LIST_MAX] = { 50.0, 95.0, 99.0, };
+
+enum {
+	HELP_LONG_OPT = 1,
+};
+
+/* this enum needs to match up with the labels array below */
+enum {
+	READ_STATS = 0,
+	WRITE_STATS,
+	ALLOC_STATS,
+	TOTAL_STATS,
+};
+
+char *stat_labels[] = {
+	"Read latency",
+	"Write latency",
+	"Allocation latency",
+	NULL,
+};
+
+char *option_string = "t:s:C:c:r:n:f:FR:T:m:W:M:w:i:D:";
+static struct option long_options[] = {
+	{"mmapsize", required_argument, 0, 'M'},
+	{"filesize", required_argument, 0, 'f'},
+	{"numfiles", required_argument, 0, 'n'},
+	{"readsize", required_argument, 0, 'R'},
+	{"writesize", required_argument, 0, 'W'},
+	{"readthreads", required_argument, 0, 'T'},
+	{"duthreads", required_argument, 0, 'D'},
+	{"threads", required_argument, 0, 't'},
+	{"runtime", required_argument, 0, 'r'},
+	{"warmuptime", required_argument, 0, 'w'},
+	{"sleeptime", required_argument, 0, 's'},
+	{"interval", required_argument, 0, 'i'},
+	{"cputime", required_argument, 0, 'c'},
+	{"cputhreads", required_argument, 0, 'C'},
+	{"memory", required_argument, 0, 'm'},
+	{"funksync", no_argument, 0, 'F'},
+	{"help", no_argument, 0, HELP_LONG_OPT},
+	{0, 0, 0, 0}
+};
+
+static void print_usage(void)
+{
+	fprintf(stderr, "simoop usage:\n"
+		"\t-t (--threads): worker threads (def: 16)\n"
+		"\t-m (--memory): memory in MB to allocate during think time in each worker (def 128)\n"
+		"\t-M (--mmapsize): amount in MB to mmap to time allocator (64MB)\n"
+		"\t-r (--runtime): How long to run before exiting (seconds, def: 30)\n"
+		"\t-w (--warmuptime): How long to warmup before resetting the stats (seconds, def: 60)\n"
+		"\t-i (--interval): Sleep time in seconds between latency reports (sec, def: 120\n"
+		"\t-s (--sleeptime): Sleep time in usecs between worker loops (usec, def: 10000\n"
+		"\t-c (--cputime): How long to think during each worker loop (seconds, def: 3)\n"
+		"\t-C (--cputhreads): How many threads do the cpu time loop (24)\n"
+		"\t-n (--numfiles): Number of files per directory tree (65536)\n"
+		"\t-f (--filesize): Size of each file in MB (64MB)\n"
+		"\t-R (--readsize): amount in MB to read from each file (2MB)\n"
+		"\t-W (--writesize): amount in MB to write to tmp files (2MB)\n"
+		"\t-T (--rwthreads): how many threads to read/write (8)\n"
+		"\t-D (--duthraeds): how many threads to scanning the working dirs (1)\n"
+		"\t-F (--funksync): should we fsync;truncate(0);fsync after writes\n"
+		"\t dir1 [dir2 ... dirN]\n"
+	       );
+	exit(1);
+}
+
+static void parse_options(int ac, char **av)
+{
+	int c;
+	int found_sleeptime = -1;
+	int found_cputime = -1;
+	int i;
+
+	while (1) {
+		int option_index = 0;
+
+		c = getopt_long(ac, av, option_string,
+				long_options, &option_index);
+
+		if (c == -1)
+			break;
+
+		switch(c) {
+		case 's':
+			found_sleeptime = atoi(optarg);
+			break;
+		case 'm':
+			thinking_mem = atoi(optarg);
+			thinking_mem *= 1024 *1024;
+			break;
+		case 'M':
+			mmap_size = atoi(optarg);
+			mmap_size *= 1024 *1024;
+			break;
+		case 'c':
+			found_cputime = atoi(optarg);
+			break;
+		case 'C':
+			cpu_threads = atoi(optarg);
+			break;
+		case 't':
+			worker_threads = atoi(optarg);
+			break;
+		case 'r':
+			runtime = atoi(optarg);
+			break;
+		case 'w':
+			warmup_seconds = atoi(optarg);
+			break;
+		case 'i':
+			interval_seconds = atoi(optarg);
+			break;
+		case 'F':
+			funksync = 0;
+			break;
+		case 'f':
+			file_size = atoi(optarg);
+			file_size *= 1024 * 1024;
+			break;
+		case 'n':
+			num_files = atoi(optarg);
+			break;
+		case 'R':
+			read_size = atoi(optarg);
+			read_size *= 1024 * 1024;
+			break;
+		case 'W':
+			write_size = atoi(optarg);
+			write_size *= 1024 * 1024;
+			break;
+		case 'T':
+			rw_threads = atoi(optarg);
+			break;
+		case 'D':
+			du_threads = atoi(optarg);
+			break;
+		case '?':
+		case HELP_LONG_OPT:
+			print_usage();
+			break;
+		default:
+			break;
+		}
+	}
+
+
+	total_paths = ac - optind;
+
+	if (total_paths <= 0) {
+		fprintf(stderr, "No directories specified\n");
+		print_usage();
+		exit(1);
+	}
+	paths = malloc(sizeof(char *) * total_paths + 1);
+	paths[total_paths] = NULL;
+	for (i = 0; i < total_paths; i++) {
+		paths[i] = strdup(av[optind++]);
+		fprintf(stderr, "adding path %s\n", paths[i]);
+	}
+
+	/*
+	 * by default pipe mode zeros out cputime and sleep time.  This
+	 * sets them to any args that were actually passed in
+	 */
+	if (found_sleeptime >= 0)
+		sleeptime = found_sleeptime;
+	if (found_cputime >= 0)
+		cputime = found_cputime * 1000000;
+
+	if (optind < ac) {
+		fprintf(stderr, "Error Extra arguments '%s'\n", av[optind]);
+		exit(1);
+	}
+}
+
+void tvsub(struct timeval * tdiff, struct timeval * t1, struct timeval * t0)
+{
+	tdiff->tv_sec = t1->tv_sec - t0->tv_sec;
+	tdiff->tv_usec = t1->tv_usec - t0->tv_usec;
+	if (tdiff->tv_usec < 0 && tdiff->tv_sec > 0) {
+		tdiff->tv_sec--;
+		tdiff->tv_usec += 1000000;
+		if (tdiff->tv_usec < 0) {
+			fprintf(stderr, "lat_fs: tvsub shows test time ran backwards!\n");
+			exit(1);
+		}
+	}
+
+	/* time shouldn't go backwards!!! */
+	if (tdiff->tv_usec < 0 || t1->tv_sec < t0->tv_sec) {
+		tdiff->tv_sec = 0;
+		tdiff->tv_usec = 0;
+	}
+}
+
+/*
+ * returns the difference between start and stop in usecs.  Negative values
+ * are turned into 0
+ */
+unsigned long long tvdelta(struct timeval *start, struct timeval *stop)
+{
+	struct timeval td;
+	unsigned long long usecs;
+
+	tvsub(&td, stop, start);
+	usecs = td.tv_sec;
+	usecs *= 1000000;
+	usecs += td.tv_usec;
+	return (usecs);
+}
+
+/* mr axboe's magic latency histogram */
+static unsigned int plat_val_to_idx(unsigned int val)
+{
+	unsigned int msb, error_bits, base, offset;
+
+	/* Find MSB starting from bit 0 */
+	if (val == 0)
+		msb = 0;
+	else
+		msb = sizeof(val)*8 - __builtin_clz(val) - 1;
+
+	/*
+	 * MSB <= (PLAT_BITS-1), cannot be rounded off. Use
+	 * all bits of the sample as index
+	 */
+	if (msb <= PLAT_BITS)
+		return val;
+
+	/* Compute the number of error bits to discard*/
+	error_bits = msb - PLAT_BITS;
+
+	/* Compute the number of buckets before the group */
+	base = (error_bits + 1) << PLAT_BITS;
+
+	/*
+	 * Discard the error bits and apply the mask to find the
+	 * index for the buckets in the group
+	 */
+	offset = (PLAT_VAL - 1) & (val >> error_bits);
+
+	/* Make sure the index does not exceed (array size - 1) */
+	return (base + offset) < (PLAT_NR - 1) ?
+		(base + offset) : (PLAT_NR - 1);
+}
+
+/*
+ * Convert the given index of the bucket array to the value
+ * represented by the bucket
+ */
+static unsigned int plat_idx_to_val(unsigned int idx)
+{
+	unsigned int error_bits, k, base;
+
+	if (idx >= PLAT_NR) {
+		fprintf(stderr, "idx %u is too large\n", idx);
+		exit(1);
+	}
+
+	/* MSB <= (PLAT_BITS-1), cannot be rounded off. Use
+	 * all bits of the sample as index */
+	if (idx < (PLAT_VAL << 1))
+		return idx;
+
+	/* Find the group and compute the minimum value of that group */
+	error_bits = (idx >> PLAT_BITS) - 1;
+	base = 1 << (error_bits + PLAT_BITS);
+
+	/* Find its bucket number of the group */
+	k = idx % PLAT_VAL;
+
+	/* Return the mean of the range of the bucket */
+	return base + ((k + 0.5) * (1 << error_bits));
+}
+
+
+static unsigned int calc_percentiles(unsigned int *io_u_plat, unsigned long nr,
+				     unsigned int **output)
+{
+	unsigned long sum = 0;
+	unsigned int len, i, j = 0;
+	unsigned int oval_len = 0;
+	unsigned int *ovals = NULL;
+	int is_last;
+
+	len = 0;
+	while (len < PLAT_LIST_MAX && plist[len] != 0.0)
+		len++;
+
+	if (!len)
+		return 0;
+
+	/*
+	 * Calculate bucket values, note down max and min values
+	 */
+	is_last = 0;
+	for (i = 0; i < PLAT_NR && !is_last; i++) {
+		sum += io_u_plat[i];
+		while (sum >= (plist[j] / 100.0 * nr)) {
+			if (j == oval_len) {
+				oval_len += 100;
+				ovals = realloc(ovals, oval_len * sizeof(unsigned int));
+			}
+
+			ovals[j] = plat_idx_to_val(i);
+			is_last = (j == len - 1);
+			if (is_last)
+				break;
+
+			j++;
+		}
+	}
+
+	*output = ovals;
+	return len;
+}
+
+static void calc_p99(struct stats *s, unsigned int *p50,
+		     unsigned int *p95, unsigned int *p99)
+{
+	unsigned int *ovals = NULL;
+	int len;
+
+	*p50 = 0;
+	*p95 = 0;
+	*p99 = 0;
+	len = calc_percentiles(s->plat, s->nr_samples, &ovals);
+	if (len && len > PLIST_P99)
+		*p99 = ovals[PLIST_P99];
+	if (len && len > PLIST_P99)
+		*p95 = ovals[PLIST_P95];
+	if (len && len > PLIST_P50)
+		*p50 = ovals[PLIST_P50];
+	if (ovals)
+		free(ovals);
+}
+
+/* fold latency info from s into d */
+void combine_stats(struct stats *d, struct stats *s)
+{
+	int i;
+	for (i = 0; i < PLAT_NR; i++)
+		d->plat[i] += s->plat[i];
+	d->nr_samples += s->nr_samples;
+	if (s->max > d->max)
+		d->max = s->max;
+	if (s->min < d->min)
+		d->min = s->min;
+}
+
+/* record a latency result into the histogram */
+static void add_lat(struct stats *s, unsigned int us)
+{
+	int lat_index = 0;
+
+	if (us > s->max)
+		s->max = us;
+	if (us < s->min)
+		s->min = us;
+
+	lat_index = plat_val_to_idx(us);
+	__sync_fetch_and_add(&s->plat[lat_index], 1);
+	__sync_fetch_and_add(&s->nr_samples, 1);
+}
+
+/*
+ * every thread has one of these, it comes out to about 19K thanks to the
+ * giant stats struct
+ */
+struct thread_data {
+	pthread_t tid;
+	/* per-thread count of worker loops over the life of the run */
+	unsigned long long work_done;
+	char *read_buf;
+	char *write_buf;
+	/* latency histogram */
+	struct stats stats[TOTAL_STATS];
+};
+
+#define nop __asm__ __volatile__("rep;nop": : :"memory")
+static void usec_spin(unsigned long spin_time)
+{
+	struct timeval now;
+	struct timeval start;
+	unsigned long long delta;
+
+	if (spin_time == 0)
+		return;
+
+	gettimeofday(&start, NULL);
+	while (1) {
+		gettimeofday(&now, NULL);
+		delta = tvdelta(&start, &now);
+		if (delta > spin_time)
+			return;
+		nop;
+	}
+}
+
+/*
+ * runs during initial file creation to create one dir
+ * in the tree
+ */
+static void make_one_dir(char *path, int a, int b)
+{
+	char subdir[256];
+	int ret;
+
+	if (b >= 0)
+		ret = snprintf(subdir, 256, "%s/%d/%d", path, a, b);
+	else
+		ret = snprintf(subdir, 256, "%s/%d", path, a);
+
+	if (ret >= 256 || ret < 0) {
+		perror("file name too long\n");
+		exit(1);
+	}
+	ret = mkdir(subdir, 0700);
+	if (ret && errno != EEXIST) {
+		perror("mkdir");
+		exit(1);
+	}
+}
+
+
+/* create the subdir tree (no files) */
+static void make_dirs(char *path)
+{
+	int first;
+	int second;
+
+	for (first = 0; first < 64; first++) {
+		make_one_dir(path, first, -1);
+		for (second = 0; second < 64; second++) {
+			make_one_dir(path, first, second);
+		}
+	}
+}
+
+/*
+ * helper to form pathnames, if postfix isn't NULL, it'll be tossed
+ * onto the end of the filename
+ */
+static void join_path(char *name, char *path, int seq, char *postfix)
+{
+	int a;
+	int b;
+	int ret;
+
+	a = seq % DIR_LEVEL;
+	b = (seq / DIR_LEVEL) % DIR_LEVEL;
+
+	if (postfix)
+		ret = snprintf(name, 256, "%s/%d/%d/%d-%s", path, a, b, seq, postfix);
+	else
+		ret = snprintf(name, 256, "%s/%d/%d/%d", path, a, b, seq);
+
+	if (ret >= 256 || ret < 0) {
+		perror("file name too long\n");
+		exit(1);
+	}
+}
+
+/* unlink working files not part of the main dataset for a given filename. */
+static void unlink_extra(char *path, int seq)
+{
+	char name[256];
+	int ret;
+
+	join_path(name, path, seq, RESULT_FILE);
+	ret = unlink(name);
+	if (ret < 0 && errno != ENOENT) {
+		perror("unlink");
+		exit(1);
+	}
+	join_path(name, path, seq, TMP_FILE);
+	ret = unlink(name);
+	if (ret < 0 && errno != ENOENT) {
+		perror("unlink");
+		exit(1);
+	}
+}
+
+/* construct a filename and return the fd */
+static int open_path(char *path, int seq, char *postfix, int flags)
+{
+	int fd;
+	char name[256];
+
+	join_path(name, path, seq, postfix);
+	fd = open(name, O_RDWR | O_CREAT | flags, 0600);
+	if (fd < 0) {
+		perror("open");
+		exit(1);
+	}
+	return fd;
+}
+
+/* helper for startup, do initial writes to a given fd */
+static void fill_one_file(int fd)
+{
+	struct stat st;
+	int ret;
+	unsigned long long cur_size;
+	char *buf;
+
+	ret = fstat(fd, &st);
+	if (ret < 0) {
+		perror("stat");
+		exit(1);
+	}
+	cur_size = st.st_size;
+
+	if (cur_size >= file_size)
+		return;
+
+	buf = malloc(BUF_SIZE);
+	if (!buf) {
+		perror("malloc");
+		exit(1);
+	}
+
+
+	memset(buf, 'a', BUF_SIZE);
+	while (cur_size < file_size) {
+		ret = write(fd, buf, BUF_SIZE);
+		if (ret < 0) {
+			perror("write");
+			exit(1);
+		}
+		if (ret < BUF_SIZE) {
+			fprintf(stderr, "short write\n");
+			exit(1);
+		}
+		cur_size += ret;
+	}
+
+	free(buf);
+}
+
+/*
+ * The du thread runs every so often and stats every single file in a
+ * given path.  This puts a lot of stress on the slab caches, and at
+ * least for XFS sets a bunch of radix bits used to track which allocation
+ * groups need to have their inodes cleaned.  It creates stress inside
+ * the shrinker.
+ */
+static void *du_thread(void *arg)
+{
+	unsigned long seq;
+	char *path = arg;
+	struct stat st;
+	int fd;
+	int ret;
+
+	while (!stopping) {
+		fprintf(stderr, "du thread is running %s\n", path);
+		for (seq = 0; seq < num_files; seq++) {
+			fd = open_path(path, seq, DATA_FILE, 0);
+			ret = fstat(fd, &st);
+			if (ret < 0 && errno != ENOENT) {
+				perror("fstat");
+				exit(1);
+			}
+			close(fd);
+		}
+		fprintf(stderr, "du thread is done %s\n", path);
+
+		/*
+		 * we need some jitter in here so all the du threads are
+		 * staggered
+		 */
+		sleep(45 + (rand() % 90));
+	}
+	return NULL;
+}
+
+/*
+ * create a temporary file and dirty it
+ */
+static void dirty_an_inode(char *path)
+{
+	int fd;
+	int seq = rand() % num_files;
+
+	fd = open_path(path, seq, TMP_FILE, 0);
+	ftruncate(fd, 100);
+	ftruncate(fd, 0);
+	close(fd);
+}
+
+static void record_one_lat(struct stats *stat, struct timeval *start,
+			   struct timeval *finish)
+{
+	unsigned long long delta;
+
+	delta = tvdelta(start, finish);
+	if (delta > 0)
+		add_lat(stat, delta);
+
+}
+
+/* reads from a random (well aligned) offset in one of the main data files */
+static void read_from_file(char *path, int seq, char *buf)
+{
+	int fd;
+	int ret;
+	int i;
+	unsigned long read_bytes = read_size;
+	unsigned long long offset;
+
+	if (read_bytes > file_size)
+		read_bytes = file_size;
+
+	/* pick a random MB starting point */
+	offset = rand() % (file_size / (1024 * 1024));
+	offset *= 1024 * 1024;
+	if (offset + read_bytes > file_size)
+		offset = file_size - read_bytes;
+
+	fd = open_path(path, seq, DATA_FILE, 0);
+
+	while (read_bytes > 0) {
+		ret = pread(fd, buf, BUF_SIZE, offset);
+		if (ret == 0)
+			break;
+		if (ret < 0) {
+			fprintf(stderr, "bad read %s seq %d ret %d offset %Lu\n", path, seq, ret, offset);
+			perror("read");
+			exit(1);
+		}
+		offset += ret;
+		read_bytes -= ret;
+	}
+
+	/* if we don't have writers making dirty inodes, make some here */
+	if (!write_size) {
+		for (i = 0; i < 8; i++)
+			dirty_an_inode(path);
+	}
+	close(fd);
+}
+
+/* creates a temp file in one of the subdirs and sends down write_bytes to it */
+static void write_to_file(char *path, int seq, char *buf)
+{
+	int fd;
+	int ret;
+	int i;
+	unsigned long write_bytes = write_size;
+	unsigned long long offset = 0;
+
+	fd = open_path(path, seq, RESULT_FILE, 0);
+
+	while (write_bytes > 0) {
+		ret = write(fd, buf, BUF_SIZE);
+		if (ret == 0)
+			break;
+		if (ret < 0) {
+			fprintf(stderr, "bad write %s seq %d ret %d offset %Lu\n", path, seq, ret, offset);
+			perror("write");
+			exit(1);
+		}
+		offset += ret;
+		write_bytes -= ret;
+	}
+	if (funksync) {
+		fsync(fd);
+		ftruncate(fd, 0);
+		fsync(fd);
+	}
+	close(fd);
+
+	/* make some dirty inodes */
+	for (i = 0; i < 8; i++)
+		dirty_an_inode(path);
+}
+
+/* make all the worker files under a main path */
+static void make_files(char *path)
+{
+	unsigned long seq;
+	int fd;
+
+	for (seq = 0; seq < num_files; seq++) {
+		fd = open_path(path, seq, DATA_FILE, 0);
+		fill_one_file(fd);
+		close(fd);
+
+		/* cleanup from the last run */
+		unlink_extra(path, seq);
+	}
+}
+
+void *filler_thread(void *arg)
+{
+	char *path = arg;
+	make_dirs(path);
+	make_files(path);
+	return 0;
+}
+
+/* start one thread per path, create the directory tree */
+void run_filler_threads(void)
+{
+	int i;
+	int ret;
+	pthread_t *tids;
+
+	tids = malloc(sizeof(*tids) * total_paths);
+	if (!tids) {
+		perror("malloc");
+		exit(1);
+	}
+	fprintf(stderr, "Creating working files\n");
+	for (i = 0; i < total_paths; i++) {
+		pthread_t tid;
+		ret = pthread_create(&tid, NULL, filler_thread,
+				     paths[i]);
+		if (ret) {
+			fprintf(stderr, "error %d from pthread_create\n", ret);
+			exit(1);
+		}
+		tids[i] = tid;
+	}
+	for (i = 0; i < total_paths; i++) {
+		pthread_join(tids[i], NULL);
+	}
+	fprintf(stderr, "done creating working files\n");
+	free(tids);
+}
+
+void *read_thread(void *arg)
+{
+	int index = rand() % total_paths;
+	int seq = rand() % num_files;
+	char *path = paths[index];
+	char *buf = arg;
+
+	read_from_file(path, seq, buf);
+	return NULL;
+}
+
+/* startup reader threads, returns the tids for later waiting */
+void read_some_files(char *buf, pthread_t *tids)
+{
+	int i;
+	int ret;
+
+	for (i = 0; i < rw_threads; i++) {
+		pthread_t tid;
+		ret = pthread_create(&tid, NULL, read_thread,
+				     buf + i * read_size);
+		if (ret) {
+			fprintf(stderr, "error %d from pthread_create\n", ret);
+			exit(1);
+		}
+		tids[i] = tid;
+	}
+}
+
+void *write_thread(void *arg)
+{
+	int index = rand() % total_paths;
+	int seq = rand() % num_files;
+	char *path = paths[index];
+	char *buf = arg;
+
+	write_to_file(path, seq, buf);
+	return NULL;
+}
+
+/* startup writer threads, returns the tids for later waiting */
+void write_some_files(char *buf, pthread_t *tids)
+{
+	int i;
+	int ret;
+
+	for (i = 0; i < rw_threads; i++) {
+		pthread_t tid;
+		ret = pthread_create(&tid, NULL, write_thread,
+				     buf + i * write_size);
+		if (ret) {
+			fprintf(stderr, "error %d from pthread_create\n", ret);
+			exit(1);
+		}
+		tids[i] = tid;
+	}
+}
+
+/* main work loop */
+void *worker_thread(void *arg)
+{
+	struct timeval now;
+	struct timeval start;
+	struct thread_data *td = arg;
+	char *read_buf;
+	char *write_buf;
+	char *mem = NULL;
+	pthread_t *read_tids;
+	pthread_t *write_tids;
+	char *mmap_ptr;
+	int i;
+	int warmup_zerod = 0;
+
+	read_tids = malloc(sizeof(*read_tids) * rw_threads);
+	write_tids = malloc(sizeof(*write_tids) * rw_threads);
+	mem = malloc(thinking_mem);
+
+	if (!read_tids || !write_tids || !mem) {
+		perror("allocation failed\n");
+		exit(1);
+	}
+
+	while(!stopping) {
+		/*
+		 * reset our stats after warmup so we don't have noise
+		 * from initial thread creation
+		 */
+		if (warmup_done && !warmup_zerod) {
+			memset(td->stats, 0, sizeof(*td->stats) * TOTAL_STATS);
+			warmup_zerod = 1;
+		}
+		read_buf = malloc(rw_threads * read_size);
+		write_buf = malloc(rw_threads * write_size);
+
+		if (!read_buf || !write_buf) {
+			perror("allocation");
+			exit(1);
+		}
+
+		/* if someone swapped out our thinking mem, bring it back */
+		memset(mem, 0, thinking_mem);
+
+		gettimeofday(&start, NULL);
+
+		/* Start the threads to read files */
+		read_some_files(read_buf, read_tids);
+
+		/* think in parallel */
+		usec_spin(cputime);
+
+		/* wait for our reads to finish */
+		for (i = 0; i < rw_threads; i++) {
+			pthread_join(read_tids[i], NULL);
+		}
+		gettimeofday(&now, NULL);
+
+		/*
+		 * record how long the reading stage took.  This
+		 * includes all of the latencies for thread creation,
+		 * doing the reads and waiting for completeion
+		 */
+		record_one_lat(&td->stats[READ_STATS], &start, &now);
+
+		/* write out the (pretend) results */
+		if (write_size) {
+			gettimeofday(&start, NULL);
+			write_some_files(write_buf, write_tids);
+			for (i = 0; i < rw_threads; i++) {
+				pthread_join(write_tids[i], NULL);
+			}
+			gettimeofday(&now, NULL);
+			record_one_lat(&td->stats[WRITE_STATS], &start, &now);
+		}
+
+		/*
+		 * we also track the latency to allocate and fault in
+		 * a chunk of pages.  This is basicaly the user-visible
+		 * impact of allocation stalls
+		 */
+		gettimeofday(&start, NULL);
+
+		mmap_ptr = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE,
+				MAP_ANONYMOUS | MAP_PRIVATE,
+				-1, 0);
+
+		if (mmap_ptr == MAP_FAILED) {
+			perror("mmap");
+			exit(1);
+		}
+
+		/* fault in all those pages */
+		for (i = 0; i < mmap_size; i += 4096) {
+			mmap_ptr[i] = 'a';
+		}
+
+		/* measure how long all of this took */
+		gettimeofday(&now, NULL);
+		record_one_lat(&td->stats[ALLOC_STATS], &start, &now);
+
+		/* think again, pretending we've done something useful */
+		munmap(mmap_ptr, mmap_size);
+		usec_spin(cputime);
+		free(read_buf);
+		free(write_buf);
+
+		td->work_done++;
+
+		usleep(sleeptime);
+	}
+
+	free(mem);
+	free(read_tids);
+	free(write_tids);
+	return NULL;
+}
+
+/*
+ * we want to keep the CPUs saturated so kswapd has to compete for CPU time
+ * these cpu threads don't do IO.
+ */
+static void *cpu_thread(void *arg)
+{
+	char *unused = arg;
+	arg = unused;
+	while(!stopping) {
+		usec_spin(cputime);
+		usleep(1);
+	}
+	return NULL;
+}
+
+/*
+ * read in /proc/vmstat so we can sum the allocation stall lines and
+ * print them out
+ */
+static int read_allocstalls(void)
+{
+	int stalls = 0;
+	int val;
+	FILE * fp;
+	char * line = NULL;
+	size_t len = 0;
+	ssize_t read;
+
+	fp = fopen("/proc/vmstat", "r");
+	if (fp == NULL)
+		return 0;
+
+	while ((read = getline(&line, &len, fp)) != -1) {
+		/*
+		 * newer kernels break out different types of allocstall,
+		 * just add them all together
+		 */
+		if (strstr(line, "allocstall")) {
+			char *p = strchr(line, ' ');
+			if (p && p[1] != '\0') {
+				val = atoi(p + 1);
+				stalls += val;
+			}
+		}
+	}
+
+	if (line)
+		free(line);
+	fclose(fp);
+	return stalls;
+}
+
+/*
+ * every worker thread tracks latencies individually.  This pulls them all
+ * into a single destination stat array for printing
+ */
+static void collect_stats(struct stats *dest, struct thread_data *worker_threads_mem)
+{
+	int i;
+	int j;
+
+	memset(dest, 0, sizeof(*dest) * TOTAL_STATS);
+	for (i = 0; i < TOTAL_STATS; i++) {
+		for (j = 0; j < worker_threads; j++)
+			combine_stats(&dest[i], &worker_threads_mem[j].stats[i]);
+	}
+	for (i = 0; i < TOTAL_STATS; i++) {
+		unsigned int p50 = 0, p95 = 0, p99 = 0;
+		calc_p99(&dest[i], &p50, &p95, &p99);
+		printf("%s (p50: %'d) (p95: %'d) (p99: %'d)\n",
+		       stat_labels[i], p50, p95, p99);
+	}
+}
+
+/*
+ * print out the current stats, along with averages and latency histogram
+ * numbers
+ */
+static void print_latencies(struct thread_data *worker_threads_mem,
+			    struct stats *stats,
+			    struct stats *work_done_stats,
+			    struct stats *allocstall_stats,
+			    double work_done, double instant_work_done,
+			    double allocstalls,
+			    double instant_allocstalls,
+			    unsigned long long delta,
+			    unsigned long long instant_delta)
+{
+	double rate;
+	double instant_rate;
+	double seconds = (double)delta / 1000000;
+	unsigned int p50, p95, p99;
+
+	printf("___\n");
+	printf("Run time: %.0f seconds\n", seconds);
+
+	/* this also prints the histogram results from the workers */
+	collect_stats(stats, worker_threads_mem);
+
+	/* calculate the work done over this period, add to histogram */
+	rate = (work_done * 1000000) / delta;
+	instant_rate = (instant_work_done * 1000000) / instant_delta;
+	add_lat(work_done_stats, rate * 100);
+	calc_p99(work_done_stats, &p50, &p95, &p99);
+
+	printf("work rate = %.2f/sec (avg %.2f/sec) (p50: %.2f) (p95: %.2f) (p99: %.2f)\n",
+	       instant_rate, rate, (double)p50/100.00, (double)p95/100.00, (double)p99/100.00);
+
+	/* do the same for the allocation stall rate */
+	rate = (allocstalls * 1000000) / delta;
+	instant_rate = (instant_allocstalls * 1000000) / instant_delta;
+	add_lat(allocstall_stats, rate * 100);
+	calc_p99(allocstall_stats, &p50, &p95, &p99);
+	printf("alloc stall rate = %.2f/sec (avg: %.2f) (p50: %.2f) (p95: %.2f) (p99: %.2f)\n",
+	       instant_rate, rate, (double)p50/100.00, (double)p95/100.00, (double)p99/100.00);
+
+}
+
+/* runtime from the command line is in seconds.  Sleep until its up */
+static void sleep_for_runtime(struct thread_data *worker_threads_mem)
+{
+	struct timeval now;
+	struct timeval start;
+	struct timeval rate_start;
+	struct timeval instant_start;
+	unsigned long long delta;
+	unsigned long long rate_delta;
+	unsigned long long instant_delta;
+	unsigned long long runtime_usec = runtime * 1000000;
+	unsigned long long warmup_usec = warmup_seconds * 1000000;
+	double work_done = 0;
+	double instant_work_done = 0;
+	double last_work_done = 0;
+	double allocstalls = 0;
+	double instant_allocstalls = 0;
+	double last_allocstalls = 0;
+	struct stats stats[TOTAL_STATS];
+	struct stats work_done_stats;
+	struct stats allocstall_stats;
+	int i;
+
+	gettimeofday(&start, NULL);
+	rate_start = start;
+
+	memset(&work_done_stats, 0, sizeof(work_done_stats));
+	memset(&allocstall_stats, 0, sizeof(allocstall_stats));
+
+	last_allocstalls = read_allocstalls();
+	allocstalls = last_allocstalls;
+	while(1) {
+		gettimeofday(&now, NULL);
+		instant_start = now;
+		delta = tvdelta(&start, &now);
+
+		if (!warmup_done && delta > warmup_usec) {
+			printf("Warmup complete (%d seconds)\n", warmup_seconds);
+			__sync_synchronize();
+			warmup_done = 1;
+			memset(&work_done_stats, 0, sizeof(work_done_stats));
+			memset(&allocstall_stats, 0, sizeof(allocstall_stats));
+			last_allocstalls = read_allocstalls();
+			last_work_done = work_done;
+			rate_start = now;
+		}
+
+		instant_allocstalls = allocstalls;
+		instant_work_done = work_done;
+		if (delta < runtime_usec)
+			sleep(interval_seconds);
+		else
+			break;
+
+		gettimeofday(&now, NULL);
+		rate_delta = tvdelta(&rate_start, &now);
+		instant_delta = tvdelta(&instant_start, &now);
+
+		work_done = 0;
+		for (i = 0; i < worker_threads; i++)
+			work_done += worker_threads_mem[i].work_done;
+
+		allocstalls = read_allocstalls();
+		if (allocstalls < last_allocstalls)
+			allocstalls = last_allocstalls;
+
+		print_latencies(worker_threads_mem, stats,
+				&work_done_stats,
+				&allocstall_stats,
+				work_done - last_work_done,
+				work_done - instant_work_done,
+				allocstalls - last_allocstalls,
+				allocstalls - instant_allocstalls,
+				rate_delta, instant_delta);
+
+	}
+	__sync_synchronize();
+	stopping = 1;
+
+	for (i = 0; i < cpu_threads; i++) {
+		pthread_join(worker_threads_mem[i + worker_threads].tid, NULL);
+	}
+
+	work_done = 0;
+	for (i = 0; i < worker_threads; i++)
+		work_done += worker_threads_mem[i].work_done;
+
+	gettimeofday(&now, NULL);
+	rate_delta = tvdelta(&rate_start, &now);
+	instant_delta = tvdelta(&instant_start, &now);
+	print_latencies(worker_threads_mem, stats,
+			&work_done_stats,
+			&allocstall_stats,
+			work_done - last_work_done,
+			work_done - instant_work_done,
+			allocstalls - last_allocstalls,
+			allocstalls - instant_allocstalls,
+			rate_delta, instant_delta);
+
+}
+
+int main(int ac, char **av)
+{
+	int i;
+	int ret;
+	int index;
+	struct thread_data *worker_threads_mem = NULL;
+	pthread_t *du_tids;
+
+	setlocale(LC_NUMERIC, "");
+	parse_options(ac, av);
+
+	if (du_threads > total_paths)
+		du_threads = total_paths;
+	du_tids = calloc(du_threads, sizeof(pthread_t));
+
+	worker_threads_mem = calloc(worker_threads + cpu_threads,
+				    sizeof(struct thread_data));
+	if (!worker_threads_mem || !du_tids) {
+		perror("calloc");
+		exit(1);
+	}
+
+	/* fill up our directory tree.  This might take a really long time */
+	run_filler_threads();
+
+	stopping = 0;
+
+	/* worker threads do the IO and the real stuff */
+	for (i = 0; i < worker_threads; i++) {
+		pthread_t tid;
+		ret = pthread_create(&tid, NULL, worker_thread,
+				     worker_threads_mem + i);
+		if (ret) {
+			perror("pthread_create");
+			exit(1);
+		}
+		worker_threads_mem[i].tid = tid;
+	}
+
+	/* CPU threads just soak up cycles */
+	for (i = 0; i < cpu_threads; i++) {
+		pthread_t tid;
+		ret = pthread_create(&tid, NULL, cpu_thread,
+				     worker_threads_mem + i + worker_threads);
+		if (ret) {
+			perror("pthread_create");
+			exit(1);
+		}
+		worker_threads_mem[i + worker_threads].tid = tid;
+	}
+
+	/*
+	 * du threads read in inodes, the goal is to have it happen on just
+	 * a couple of paths
+	 */
+	index = rand();
+	for (i = 0; i < du_threads; i++) {
+		ret = pthread_create(&du_tids[i], NULL, du_thread,
+				     paths[index++ % total_paths]);
+		if (ret) {
+			fprintf(stderr, "error %d from pthread_create\n", ret);
+			exit(1);
+		}
+	}
+
+	/* let all the magic happen and collect results */
+	sleep_for_runtime(worker_threads_mem);
+
+	for (i = 0; i < du_threads; i++) {
+		pthread_join(du_tids[i], NULL);
+	}
+
+	free(worker_threads_mem);
+	return 0;
+}