From: Waldemar Kozaczuk <jwkozac...@gmail.com>
Committer: Waldemar Kozaczuk <jwkozac...@gmail.com>
Branch: master
futex: add benchmark test
This patch adds a new benchmark - misc-futex-perf.cc. The goal is to
indirectly measure performance of the futex syscall implemented in OSv
and compare it to Linux and thus guide us in implementing the improvements
described in the issue #853.
The benchmark program does it by implementing special mutex - fmutex -
based on futex syscall according to the algorithm specified in the
Ulrich Drepper's paper "Futexes Are Tricky". The test is similar to the
misc-mutex2.cc written by Nadav Har'El. It takes three parameters:
mandatory number of threads (nthreads) and a computation length (worklen)
and optional number of mutexes (nmutexes) which is equal to 1 by default.
The test groups all threads (nthreads * nmutexes) into nmutexes sets
of nthreads threads trying to take the group mutex (one out of nmutexes)
in a loop and increment the group counter and then do some short computation
of the specified length outside the loop. The test runs for 30 seconds, and
shows the average total number of lock-protected counter increments per second.
The number of cpus is set by using the '-c' option passes to run.py in
case of OSv, and using taskset -c 0..n when running the same program on
host.
The results of the test that show number of total increments (across
counters of all groups of threads) per second for both OSv and Linux
host are below. It also shows number of total futex syscall calls (wake)
captured by adding an atomic counter in the futex implementation for OSv.
+------------------+----------------------------+----------------------+
| Run parameters | On OSv guest | On Linux host (op/s) |
| | (op/s) (futex called) |
+------------------+----------------------------+----------------------+
| 1 0 1 (1 cpu) | 5.1353e+07 0 | 5.21169e+07 |
| 2 0 1 (2 cpus) | 2.26067e+07 345,745 | 1.78575e+07 |
| 4 0 1 (4 cpus) | 4.93204e+07 2342 | 1.41494e+07 |
| 1 500 1 (1 cpu) | 5.67558e+06 0 | 5.75555e+06 |
| 2 500 1 (2 cpus) | 9.19294e+06 3618 | 9.78263e+06 |
| 4 500 1 (4 cpus) | 5.65933e+06 38,243 | 6.87465e+06 |
| 4 500 2 (4 cpus) | 8.30834e+06 266 | 1.15537e+07 |
| 4 500 4 (4 cpus) | 1.06216e+07 111 | 1.16908e+07 |
| 4 500 8 (4 cpus) | 1.39291e+07 101 | 1.31845e+07 |
+------------------+----------------------------+----------------------+
The results are surprising and somewhat confusing. For example the lines
2 and 3 show OSv outperforming Linux by a lot. Also the line 7 (4 500 2)
shows OSv peformance worse by ~30% even when number of futex calls is
pretty low. Possibly there is a flaw in this test, or some kind of
different explanation.
Signed-off-by: Waldemar Kozaczuk <jwkozac...@gmail.com>
Message-Id: <20220907032208.20291-1-jwkozac...@gmail.com>
---
diff --git a/modules/tests/Makefile b/modules/tests/Makefile
--- a/modules/tests/Makefile
+++ b/modules/tests/Makefile
@@ -134,7 +134,8 @@ tests := tst-pthread.so misc-ramdisk.so tst-vblk.so
tst-bsd-evh.so \
tst-elf-init.so tst-realloc.so tst-setjmp.so \
libtls.so libtls_gold.so tst-tls.so tst-tls-gold.so tst-tls-pie.so \
tst-sigaction.so tst-syscall.so tst-ifaddrs.so tst-getdents.so \
- tst-netlink.so misc-zfs-io.so misc-zfs-arc.so tst-pthread-create.so
+ tst-netlink.so misc-zfs-io.so misc-zfs-arc.so tst-pthread-create.so \
+ misc-futex-perf.so
# libstatic-thread-variable.so tst-static-thread-variable.so \
ifeq ($(arch),x64)
diff --git a/tests/misc-futex-perf.cc b/tests/misc-futex-perf.cc
--- a/tests/misc-futex-perf.cc
+++ b/tests/misc-futex-perf.cc
@@ -0,0 +1,190 @@
+/*
+ * Copyright (C) 2022 Waldemar Kozaczuk
+ *
+ * This work is open source software, licensed under the terms of the
+ * BSD license as described in the LICENSE file in the top-level directory.
+ */
+#include <stdint.h>
+#include <unistd.h>
+#include <sys/syscall.h>
+#include <sys/sysinfo.h>
+#include <linux/futex.h>
+#include <thread>
+#include <chrono>
+#include <iostream>
+#include <vector>
+
+// This test is based on misc-mutex2.cc written by Nadav Har'El. But unlike
+// the other one, it focuses on measuring the performance of the futex()
+// syscall implementation. It does it indirectly by implementing mutex based
+// on futex syscall according to the formula specified in the Ulrich Drepper's
+// paper "Futexes Are Tricky".
+// It takes three parameters: mandatory number of threads (nthreads) and
+// a computation length (worklen) and optional number of mutexes (nmutexes).
+// The test groups all threads (nthreads * nmutexes) into nmutexes sets
+// where nthreads threads loop trying to take the group mutex (one out of
nmutexes)
+// and increment the group counter and then do some short computation of the
+// specified length outside the loop. The test runs for 30 seconds, and
+// shows the average number of lock-protected counter increments per second.
+// The reason for doing some computation outside the lock is that makes the
+// benchmark more realistic, reduces the level of contention and makes it
+// beneficial for the OS to run the different threads on different CPUs:
+// Without any computation outside the lock, the best performance will be
+// achieved by running all the threads.
+
+// Turn off optimization, as otherwise the compiler will optimize
+// out calls to fmutex lock() and unlock() as they seem to do nothing
+#pragma GCC optimize("00")
+
+// Wrapper function that performs the same functionality as described
+// in the Drepper's paper (see below).
+// It atomically compares the value pointed by the address addr to the value
expected
+// and only if equal replaces *addr with desired. In either case it returns
the value
+// at *addr before the operation.
+inline uint32_t cmpxchg(uint32_t *addr, uint32_t expected, uint32_t desired)
+{
+ uint32_t *expected_addr = &expected;
+ __atomic_compare_exchange_n(addr, expected_addr, desired, false,
__ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST);
+ return *expected_addr;
+}
+
+enum {
+ UNLOCKED = 0,
+ LOCKED_NO_WAITERS = 1,
+ LOCKED_MAYBE_WAITERS = 2,
+};
+
+// This futex-based mutex implementation is based on the example "Mutex, Take
2"
+// from the Ulrich Drepper's paper "Futexes Are Tricky"
(https://dept-info.labri.fr/~denis/Enseignement/2008-IR/Articles/01-futex.pdf)
+class fmutex {
+public:
+ fmutex() : _state(UNLOCKED) {}
+ void lock()
+ {
+ uint32_t c;
+ // If the state was UNLOCKED before cmpxchg, we do not have to do
anything
+ // just return after setting to LOCKED_NO_WAITERS
+ if ((c = cmpxchg(&_state, UNLOCKED, LOCKED_NO_WAITERS)) != UNLOCKED) {
+ do {
+ // It was locked, so let us set the state to
LOCKED_MAYBE_WAITERS.
+ // It might be already in this state (1st part of if below) or
+ // the state was LOCKED_NO_WAITERS so let us change it to
LOCKED_MAYBE_WAITERS
+ if (c == LOCKED_MAYBE_WAITERS ||
+ cmpxchg(&_state, LOCKED_NO_WAITERS, LOCKED_MAYBE_WAITERS)
!= UNLOCKED) {
+ // Wait until kernel tells us the state is different from
LOCKED_MAYBE_WAITERS
+ syscall(SYS_futex, &_state, FUTEX_WAIT_PRIVATE,
LOCKED_MAYBE_WAITERS, 0, 0, 0);
+ }
+ // At this point we are either because:
+ // 1. The mutex was indeed UNLOCKED = the if condition above
was false
+ // 2. We were awoken when sleeping upon making the syscall
FUTEX_WAIT_PRIVATE
+ // So let us try to lock again. Because we do not know if
there any waiters
+ // we try to set to LOCKED_MAYBE_WAITERS and err on the safe
side.
+ } while ((c = cmpxchg(&_state, UNLOCKED, LOCKED_MAYBE_WAITERS)) !=
UNLOCKED);
+ }
+ }
+
+ void unlock()
+ {
+ // Let us wake one waiter only if the state was LOCKED_MAYBE_WAITERS
+ // Otherwise do nothing if uncontended
+ if (__atomic_fetch_sub(&_state, 1, __ATOMIC_SEQ_CST) !=
LOCKED_NO_WAITERS) {
+ _state = UNLOCKED;
+ syscall(SYS_futex, &_state, FUTEX_WAKE_PRIVATE, 1, 0, 0, 0);
+ }
+ }
+private:
+ uint32_t _state;
+};
+
+void loop(int iterations)
+{
+ for (register int i=0; i<iterations; i++) {
+ // To force gcc to not optimize this loop away
+ asm volatile("" : : : "memory");
+ }
+}
+
+int main(int argc, char** argv) {
+ if (argc <= 2) {
+ std::cerr << "Usage: " << argv[0] << " nthreads worklen <nmutexes>\n";
+ return 1;
+ }
+ int nthreads = atoi(argv[1]);
+ if (nthreads <= 0) {
+ std::cerr << "Usage: " << argv[0] << " nthreads worklen <nmutexes>\n";
+ return 2;
+ }
+ // "worklen" is the amount of work to do in each loop iteration, outside
+ // the mutex. This reduces contention and makes the benchmark more
+ // realistic and gives it the theoretic possibility of achieving better
+ // benchmark numbers on multiple CPUs (because this "work" is done in
+ // parallel.
+ int worklen = atoi(argv[2]);
+ if (worklen < 0) {
+ std::cerr << "Usage: " << argv[0] << " nthreads worklen <nmutexes>\n";
+ return 3;
+ }
+
+ // "nmutexes" is the number of mutexes the threads will be contending for
+ // we will group threads by set of nthreads contending on individual mutex
+ // to increase corresponding group counter
+ int nmutexes = 1;
+ if (argc >= 4) {
+ nmutexes = atoi(argv[3]);
+ if (nmutexes < 0)
+ nmutexes = 1;
+ }
+
+ int concurrency = 0;
+ cpu_set_t cs;
+ sched_getaffinity(0, sizeof(cs), &cs);
+ for (int i = 0; i < get_nprocs(); i++) {
+ if (CPU_ISSET(i, &cs)) {
+ concurrency++;
+ }
+ }
+ std::cerr << "Running " << (nthreads * nmutexes) << " threads on " <<
+ concurrency << " cores with " << nmutexes <<
+ " mutexes. Worklen = " <<
+ worklen << "\n";
+
+ // Set secs to the desired number of seconds a measurement should
+ // take. Note that the whole test will take several times longer than
+ // secs, as we do several tests each lasting at least this long.
+ double secs = 30.0;
+
+ // Our mutex-protected operation will be a silly increment of a counter,
+ // taking a tiny amount of time, but still can happen concurrently if
+ // run very frequently from many cores in parallel.
+ long counters[nmutexes] = {0};
+ bool done = false;
+
+ fmutex mut[nmutexes];
+ std::vector<std::thread> threads;
+ for (int m = 0; m < nmutexes; m++) {
+ for (int i = 0; i < nthreads; i++) {
+ threads.push_back(std::thread([&, m]() {
+ while (!done) {
+ mut[m].lock();
+ counters[m]++;
+ mut[m].unlock();
+ loop(worklen);
+ }
+ }));
+ }
+ }
+ threads.push_back(std::thread([&]() {
+ std::this_thread::sleep_for(std::chrono::duration<double>(secs));
+ done = true;
+ }));
+ for (auto &t : threads) {
+ t.join();
+ }
+ long total = 0;
+ for (int m = 0; m < nmutexes; m++) {
+ total += counters[m];
+ }
+ std::cout << total << " counted in " << secs << " seconds (" <<
(total/secs) << " per sec)\n";
+
+ return 0;
+}
--
You received this message because you are subscribed to the Google Groups "OSv
Development" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to osv-dev+unsubscr...@googlegroups.com.
To view this discussion on the web visit
https://groups.google.com/d/msgid/osv-dev/000000000000e460af05eb296bd9%40google.com.
