Futex Scaling for Multi-core Systems

Futex Scaling for Multicore Systems
ACM Applicative Conference – June 2016. New York, NY.
Davidlohr Bueso <dave@stgolabs.net>
SUSE Labs.

2
Agenda
1. Introduction
2. Implementing Futexes
● Overall architecture.
● Addressing performance bottlenecks.
3. Notes/Practises in Userspace.

3
Introduction
• Linux kernel (v2.5) functionality for userspace: “Fast
userspace mutual exclusion” through the futex(2)
interface:
‒ Method for a program to wait for a value at a given address to
change, and a method to wake up anyone waiting on a particular
address.
‒ A futex is in essence a userspace address.

4
Introduction
• Linux kernel (v2.5) functionality for userspace: “Fast
userspace mutual exclusion” through the futex(2)
interface:
‒ Method for a program to wait for a value at a given address to
change, and a method to wake up anyone waiting on a particular
address.
‒ A futex is in essence a userspace address.
• Futexes are very basic and lend themselves well for
building higher level locking abstractions such as POSIX
threads:
‒ pthread_mutex_*(), pthread_rwlock_*(),
pthread_barrier_*(), pthread_cond_wait(), etc.

5
Introduction
• In the uncontended cases, user locking implementations
never need to exit from userspace, and the kernel is
graciously unaware, nor cares. CAS is enough.

6
Introduction
• In the uncontended cases, user locking implementations
never need to exit from userspace, and the kernel is
graciously unaware, nor cares. CAS is enough.
• In the case of sysv sems this is not true as jumping to
kernel space is always required to handle the call.
• Lock fastpaths therefore have a significant advantage
using by futexes.

7
Introduction
int futex(int *uaddr, int futex_op,
int val, struct timespec *to,
int *uaddr2, int val3);

8
Introduction
‒ The futex, 32-bit lock variable field

9
Introduction
‒ What operation to do on the futex, ie:
FUTEX_WAIT, FUTEX_WAKE

10
Introduction
‒ What operation to do on the futex, ie:
FUTEX_WAIT, FUTEX_WAKE, FUTEX_REQUEUE, etc.

11
Introduction
• Special cases (operations):
‒ PI-futexes (PTHREAD_PRIO_INHERIT)
FUTEX_LOCK/UNLOCK_PI
FUTEX_CMP/WAIT_REQUEUE_PI, etc.

12
Introduction
• Special cases (operations):
‒ PI-futexes (PTHREAD_PRIO_INHERIT)
FUTEX_LOCK/UNLOCK_PI
FUTEX_CMP/WAIT_REQUEUE_PI, etc.
‒ Robust futexes (lock owner crashes)
set_robust_list(2), get_robust_list(2)

13
Introduction
void lock(uint32_t *futex)
{
        uint32_t old = *futex;
        if (old == UNLOCKED) /* fastpath */
                old = cmpxchg(futex, UNLOCKED, LOCKED);
        while (old != UNLOCKED) {
                if (old == LOCKED)
                        old = cmpxchg(futex, LOCKED, WAITER);
                if (old != UNLOCKED) {
                        futex(futex, FUTEX_WAIT, WAITERS,…);
                        old = *futex;
                }
                if (old == UNLOCKED)
                        old = cmpxchg(futex, UNLOCKED, WAITERS);
        }
}

14
Introduction
void unlock(uint32_t *futex)
{
        uint32_t old = *futex;
        if (old == LOCKED)
                old = cmpxchg(futex, LOCKED, UNLOCK);
        if (old == WAITER) {
                old = cmpxchg(futex, WAITER, UNLOCKED);
                nwakes = futex(futex, FUTEX_WAKE, 1, ...);
    /* check nwakes == 1 */
        }
}

16
Kernel Implementation
• The uaddr is used by the kernel to create a unique futex
key, each key hashes to a hash bucket.
• The task’s stack holds the futex_q chain when waiting
(servicing FUTEX_WAIT operations).

17
• Wait queues are at the heart of futexes.
‒ Priority queues (high prio tasks first, otherwise FIFO).
‒ Governed by a chained global hash table.

18
• Each bucket is serialized by a spinlock – all operations
require holding the lock beforehand.
• One or more futexes can share the queue (collisions).

19
Bottlenecks
• There are some immediately apparent issues with the
current futex architecture.
‒ Global hash table (really bad for NUMA).
‒ Hash table collisions.
‒ hb>lock contention/hold times.

20
Bottlenecks
• All of these can have disastrous effects on both
performance, as systems increase in hardware
capabilities, as well as determinism for real-time.

21
Bottlenecks
• All of these can have disastrous effects on both
performance, as systems increase in hardware
capabilities, as well as determinism for real-time.
• Numerous efforts have been taken to mitigate some of
these scalability problems.

22
Keys and Hashing
• Uses Jenkins hash function (lookup3).
‒ Fast and distributes hash values rather uniformly
(on real workloads).
• Keys for private vs shared futexes.
‒ Private simply use the current address space and the futex
uaddr.
‒ Shared mappings require page pinning (gup), locks, RCU, ref
counting, etc. Even worse if inode-backed.
• For shared mappings, lockless get_futex_key()
‒ Avoids taking the page_lock (sleepable).
‒ Good for performance and RT.

23
Keys and Hashing
52-core, 2 socket x86-64 (Haswell)
nfutexes = nthreads * 1024

24
Keys and Hashing
• Avoiding collisions and therefore improving the
parallelisation of different futexes is a major plus.
‒ Ie: two or more user locks can be operated on concurrently
without being serialized by the same hb>lock.
‒ The perfect hash size will of course have one to one hb:futex
ratio.

25
Keys and Hashing
• Futexes started out at 256 entry hash table, which
caused havoc on multicore systems. Since then we
scale by number of CPUs (and avoid false sharing).
‒ Improved raw hashing throughput by 80% to 800% in
increasing futex counts.

26
Per-process Hash Table
• Recent patchset proposed upstream to address the
NUMA issues of the global table for private futexes.
• Dynamically sized: if a potential collision is detected the
size of the hash table is doubled.
• Hash table being on the same NUMA node as the task
operating on the futex.
• Addresses collisions by dedicating more hash table space
per process.

28
Hash Bucket Lock Contention
• For a successful futex call to occur, intuitively, among
others, the following work must occur while holding
the the hb>lock
‒ Priority list handling.
‒ Block/wakeup(s).
• It is not hard to find pathological contention on some
hb>lock, when multiple operations are being done
on the same futex/lock.

29
Lockless Wakeups
• Internally acknowledge that one or more tasks are to
be awoken, then call wake_up_process() after
releasing the bucket spinlock.

30
Lockless Wakeups
• Internally acknowledge that one or more tasks are to
be awoken, then call wake_up_process() after
releasing the bucket spinlock.
• Lockless wake-queues respect the order given by the
caller, hence wakeup fairness does not change
whatsoever.

31
Lockless Wakeups
Works particularly well for batch wakeups of tasks
blocked on a particular futex.
‒ Ie waking all reader-waiters that where blocked on some lock
held for write. (Where N is a large number):
futex(uaddr, FUTEX_WAKE, N, ...);

32
Lockless Wakeups
16
32
48
64
80
96
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08
parallel-wakeups
no-batch
batched
time (ms)
threads
26-core (HT), 2 socket x86-64 (Haswell)

33
Queued/MCS Spinlocks (x86)
• Bottlenecks in userspace can easily lead to severe
contention on the hb>lock, and therefore exposed
to the semantics of spinlocks.

35
• Replaced the regular ticket spinlock implementation.
• Each lock waiter will be queued and spins on its own
cacheline (per-cpu variable) rather than the lock itself.
‒ This occurs until the waiter becomes the head of the queue
(next in line to take the lock).
‒ Eliminates much of the cacheline bouncing (inter-socket traffic)
caused by contended ticket locks.

36
• Replaced the regular ticket spinlock implementation.
• Each lock waiter will be queued and spins on its own
cacheline (per-cpu variable) rather than the lock itself.
‒ This occurs until the waiter becomes the head of the queue
(next in line to take the lock).
‒ Eliminates much of the cacheline bouncing (inter-socket traffic)
caused by contended ticket locks.
• This really matters on systems with > 4-sockets, but
can bring 8 or 16-socket machines to its knees.
‒ Experiments show improvements in throughput of up to 2.4x
on 80 core machines.
‒ Reports of lockups for futexes on 240-core systems.

37
• qspinlocks outperform ticket locks in the uncontended
case. Ie avg single threaded lock+unlock:
• Therefore smaller systems under non-pathological
(normal case) workloads can also benefit.
Time (ns)
Ticket lock (unlock: CAS) 17.63
Queued lock (unlock: store) 9.54
(2.6Ghz x86-64)

38
PI-Futexes
• Futexes make use of rt-mutexes to support priority-
inheritance (PTHREAD_PRIO_INHERIT) semantics.
‒ pi_state is attached to the waiter’s futex_q
‒ The pi_state>pi_mutex top-waiter (highest priority
waiter) has been optimized for both lockless wakeups and
avoid blocking if current lock owner is running.

39
PI-Futexes
1 4 8 16 24 48 64
0
2000000
4000000
6000000
8000000
10000000
12000000
14000000
pistress benchmark
top-waiter
regular
groups of 3-threads
totalinversions
32-core, 2 socket x86-64

41
General Notes
• The performance optimizations at the Kernel side are only
one part of the picture. Using futexes is just as important.
• As with any system call, there really is no single recipe to
make good use of futexes in userspace. The kernel simply
obliges.
• Locking algorithms can play a huge factor in performance
on large-scale machines.
‒ Contention on a 240-core system is much more severe than
on a 40-core machine.

42
General Notes
• Locks in both the kernel and in userspace can be exposed
to the same architectural difficulties: cacheline contention
and NUMA-awareness.
• Many applications today are developed/tuned for certain
amount of CPUs.
‒ Scaling based only on the number of CPUs is likely to introduce
significant lock and cacheline contention.
• Unsurprisingly similar optimizations and tools to obtain
data for analysis (perf, tracing, etc) can be taken from this
presentation and applied to your locks.

43
Best Practises
• Data partitioning.
‒ Cacheline contention within a single NUMA node can be
significantly less severe than among cores from different NUMA
nodes.
• Lock granularity.
• Data layout
‒ structure organization, avoiding false sharing.
‒ Cacheline bouncing can occur when there are multiple hb>lock
residing on the same cacheline and different futexes hash to
adjacent buckets.
• Avoid futex(2) calls unless necessary
‒ Ie: make sure there are waiters to wakeup.

44
References
• man 2 futex
• Hart, Darren. “A futex overview and update”. lwn.net. Nov 2009.
• Drepper, Ulrich. “Futexes are Tricky”. Nov 2011.
• Hart, D. “Requeue-PI: Making Glibc Condvars PI-Aware”. Proc. RT
Linux Summit 2011.
• Bueso, D. Norton, S. “An Overview of Kernel Lock Improvements”.
Linux Con. 2014. Chicago, IL.

Futex Scaling for Multi-core Systems

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Futex Scaling for Multi-core Systems (20)

Recently uploaded (20)

Futex Scaling for Multi-core Systems