From: Lihao Liang <lihaoliang@google.com>
To: Alex Kogan <alex.kogan@oracle.com>
Cc: linux-arch@vger.kernel.org, guohanjun@huawei.com, arnd@arndb.de,
Peter Zijlstra <peterz@infradead.org>,
dave.dice@oracle.com, jglauber@marvell.com, x86@kernel.org,
will.deacon@arm.com, linux@armlinux.org.uk,
steven.sistare@oracle.com, linux-kernel@vger.kernel.org,
mingo@redhat.com, bp@alien8.de, hpa@zytor.com,
longman@redhat.com, tglx@linutronix.de,
daniel.m.jordan@oracle.com, Will Deacon <will@kernel.org>,
linux-arm-kernel@lists.infradead.org
Subject: Re: [PATCH v9 0/5] Add NUMA-awareness to qspinlock
Date: Sun, 26 Jan 2020 01:58:02 +0000 [thread overview]
Message-ID: <CAC4j=Y--5UQR7Oc5n+sxAwLxd_PKi0Eb-7aiZjDTUW0FTJy8Tw@mail.gmail.com> (raw)
In-Reply-To: <CAC4j=Y_SMHe4WNpaaS0kK1JYfnufM+AAiwwUMBx27L8U6bD8Yg@mail.gmail.com>
On Sun, Jan 26, 2020 at 12:32 AM Lihao Liang <lihaoliang@google.com> wrote:
>
> Hi Alex and Waiman,
>
> Thanks a lot for your swift response and clarification.
>
> On Wed, Jan 22, 2020 at 7:30 PM Alex Kogan <alex.kogan@oracle.com> wrote:
> >
> > Hi, Lihao.
> >
> > > On Jan 22, 2020, at 6:45 AM, Lihao Liang <lihaoliang@google.com> wrote:
> > >
> > > Hi Alex,
> > >
> > > On Wed, Jan 22, 2020 at 10:28 AM Alex Kogan <alex.kogan@oracle.com> wrote:
> > >>
> > >> Summary
> > >> -------
> > >>
> > >> Lock throughput can be increased by handing a lock to a waiter on the
> > >> same NUMA node as the lock holder, provided care is taken to avoid
> > >> starvation of waiters on other NUMA nodes. This patch introduces CNA
> > >> (compact NUMA-aware lock) as the slow path for qspinlock. It is
> > >> enabled through a configuration option (NUMA_AWARE_SPINLOCKS).
> > >>
> > >
> > > Thanks for your patches. The experimental results look promising!
> > >
> > > I understand that the new CNA qspinlock uses randomization to achieve
> > > long-term fairness, and provides the numa_spinlock_threshold parameter
> > > for users to tune.
> > This has been the case in the first versions of the series, but is not true anymore.
> > That is, the long-term fairness is achieved deterministically (and you are correct
> > that it is done through the numa_spinlock_threshold parameter).
> >
> > > As Linux runs extremely diverse workloads, it is not
> > > clear how randomization affects its fairness, and how users with
> > > different requirements are supposed to tune this parameter.
> > >
> > > To this end, Will and I consider it beneficial to be able to answer the
> > > following question:
> > >
> > > With different values of numa_spinlock_threshold and
> > > SHUFFLE_REDUCTION_PROB_ARG, how long do threads running on different
> > > sockets have to wait to acquire the lock?
> > The SHUFFLE_REDUCTION_PROB_ARG parameter is intended for performance
> > optimization only, and *does not* affect the long-term fairness (or, at the
> > very least, does not make it any worse). As Longman correctly pointed out in
> > his response to this email, the shuffle reduction optimization is relevant only
> > when the secondary queue is empty. In that case, CNA hands-off the lock
> > exactly as MCS does, i.e., in the FIFO order. Note that when the secondary
> > queue is not empty, we do not call probably().
> >
> > > This is particularly relevant
> > > in high contention situations when new threads keep arriving on the same
> > > socket as the lock holder.
> > In this case, the lock will stay on the same NUMA node/socket for
> > 2^numa_spinlock_threshold times, which is the worst case scenario if we
> > consider the long-term fairness. And if we have multiple nodes, it will take
> > up to 2^numa_spinlock_threshold X (nr_nodes - 1) + nr_cpus_per_node
> > lock transitions until any given thread will acquire the lock
> > (assuming 2^numa_spinlock_threshold > nr_cpus_per_node).
> >
>
> You're right that the latest version of the patch handles long-term fairness
> deterministically.
>
> As I understand it, the n-th thread in the main queue is guaranteed to
> acquire the lock after N lock handovers, where N is bounded by
>
> n - 1 + 2^numa_spinlock_threshold * (nr_nodes - 1)
>
> I'm not sure what role the variable nr_cpus_per_node plays in your analysis.
>
> Do I miss anything?
>
If I understand correctly, there are two phases in the algorithm:
MCS phase: when the secondary queue is empty, as explained in your emails,
the algorithm hands the lock to threads in the main queue in an FIFO order.
When probably(SHUFFLE_REDUCTION_PROB_ARG) returns false (with default
probability 1%), if the algorithm finds the first thread running on the same
socket as the lock holder in cna_scan_main_queue(), it enters the following
CNA phase.
CNA phase: when the secondary queue is not empty, the algorithm keeps
handing the lock to threads in the main queue that run on the same socket as
the lock holder. When 2^numa_spinlock_threshold is reached, it splices
the secondary queue to the front of the main queue. And we are back to the
MCS phase above.
For the n-th thread T in the main queue, the MCS phase handles threads that
arrived in the main queue before T. In high contention situations, the CNA
phase handles two kinds of threads:
1. Threads ahead of T that run on the same socket as the lock holder when
a transition from the MCS to CNA phase was made. Assume there are m such
threads.
2. Threads that keep arriving on the same socket as the lock holder. There
are at most 2^numa_spinlock_threshold of them.
Then the number of lock handovers in the CNA phase is max(m,
2^numa_spinlock_threshold). So the total number of lock handovers before T
acquires the lock is at most
n - 1 + 2^numa_spinlock_threshold * (nr_nodes - 1)
Please let me know if I misunderstand anything.
Many thanks,
Lihao.
_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
next prev parent reply other threads:[~2020-01-26 1:58 UTC|newest]
Thread overview: 39+ messages / expand[flat|nested] mbox.gz Atom feed top
2020-01-15 3:59 [PATCH v9 0/5] Add NUMA-awareness to qspinlock Alex Kogan
2020-01-15 3:59 ` [PATCH v9 1/5] locking/qspinlock: Rename mcs lock/unlock macros and make them more generic Alex Kogan
2020-01-15 3:59 ` [PATCH v9 2/5] locking/qspinlock: Refactor the qspinlock slow path Alex Kogan
2020-01-15 3:59 ` [PATCH v9 3/5] locking/qspinlock: Introduce CNA into the slow path of qspinlock Alex Kogan
2020-01-23 9:26 ` Peter Zijlstra
2020-01-23 10:06 ` Peter Zijlstra
2020-01-23 10:16 ` Peter Zijlstra
2020-01-23 11:22 ` Will Deacon
2020-01-23 13:17 ` Peter Zijlstra
2020-01-23 14:15 ` Waiman Long
2020-01-23 15:29 ` Peter Zijlstra
2020-01-15 3:59 ` [PATCH v9 4/5] locking/qspinlock: Introduce starvation avoidance into CNA Alex Kogan
2020-01-23 19:55 ` Waiman Long
2020-01-23 20:39 ` Waiman Long
2020-01-23 23:39 ` Alex Kogan
2020-01-15 3:59 ` [PATCH v9 5/5] locking/qspinlock: Introduce the shuffle reduction optimization " Alex Kogan
2020-01-22 11:45 ` [PATCH v9 0/5] Add NUMA-awareness to qspinlock Lihao Liang
2020-01-22 17:24 ` Waiman Long
2020-01-23 11:35 ` Will Deacon
2020-01-23 15:25 ` Waiman Long
2020-01-23 19:08 ` Waiman Long
2020-01-22 19:29 ` Alex Kogan
2020-01-26 0:32 ` Lihao Liang
2020-01-26 1:58 ` Lihao Liang [this message]
2020-01-27 16:01 ` Alex Kogan
2020-01-29 1:39 ` Lihao Liang
2020-01-27 6:16 ` Alex Kogan
2020-01-24 22:24 ` Paul E. McKenney
[not found] ` <6AAE7FC6-F5DE-4067-8BC4-77F27948CD09@oracle.com>
2020-01-25 0:57 ` Paul E. McKenney
2020-01-25 1:59 ` Waiman Long
[not found] ` <adb4fb09-f374-4d64-096b-ba9ad8b35fd5@redhat.com>
2020-01-25 4:58 ` Paul E. McKenney
2020-01-25 19:41 ` Waiman Long
2020-01-26 15:35 ` Paul E. McKenney
2020-01-26 22:42 ` Paul E. McKenney
2020-01-26 23:32 ` Paul E. McKenney
2020-01-27 6:04 ` Alex Kogan
2020-01-27 14:11 ` Waiman Long
2020-01-27 15:09 ` Paul E. McKenney
[not found] ` <9b3a3f16-5405-b6d1-d023-b85f4aab46dd@redhat.com>
2020-01-27 17:17 ` Waiman Long
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to='CAC4j=Y--5UQR7Oc5n+sxAwLxd_PKi0Eb-7aiZjDTUW0FTJy8Tw@mail.gmail.com' \
--to=lihaoliang@google.com \
--cc=alex.kogan@oracle.com \
--cc=arnd@arndb.de \
--cc=bp@alien8.de \
--cc=daniel.m.jordan@oracle.com \
--cc=dave.dice@oracle.com \
--cc=guohanjun@huawei.com \
--cc=hpa@zytor.com \
--cc=jglauber@marvell.com \
--cc=linux-arch@vger.kernel.org \
--cc=linux-arm-kernel@lists.infradead.org \
--cc=linux-kernel@vger.kernel.org \
--cc=linux@armlinux.org.uk \
--cc=longman@redhat.com \
--cc=mingo@redhat.com \
--cc=peterz@infradead.org \
--cc=steven.sistare@oracle.com \
--cc=tglx@linutronix.de \
--cc=will.deacon@arm.com \
--cc=will@kernel.org \
--cc=x86@kernel.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).