Skip site navigation (1)Skip section navigation (2)
Date:      Wed, 26 Sep 2007 08:59:37 +0200
From:      "Attilio Rao" <attilio@freebsd.org>
To:        "Jeff Roberson" <jroberson@chesapeake.net>
Cc:        freebsd-smp@freebsd.org, freebsd-arch@freebsd.org
Subject:   Re: rwlocks: poor performance with adaptive spinning
Message-ID:  <3bbf2fe10709252359h7812ff7bsce9d393f80b65c06@mail.gmail.com>
In-Reply-To: <20070925155504.W556@10.0.0.1>
References:  <3bbf2fe10709221932i386f65b9h6f47ab4bee08c528@mail.gmail.com> <200709241152.41660.jhb@freebsd.org> <20070924135554.F547@10.0.0.1> <200709251314.42707.jhb@freebsd.org> <20070925155504.W556@10.0.0.1>

next in thread | previous in thread | raw e-mail | index | archive | help
2007/9/26, Jeff Roberson <jroberson@chesapeake.net>:
> On Tue, 25 Sep 2007, John Baldwin wrote:
>
> >
> > So I looked at this some more and now I don't understand why the trywait() and
> > cancel() changes were done.  Some background:
>
> I wanted the turnstile chain lock to protect only the association between
> the turnstile and lock.  This way it further reduces the contention domain
> for the hard lock/unlock cases.  We sometimes see a lot of contention on
> the turnstile chain that may be due to hash collisions.  If this reduced
> contention doesn't help I think it would be possible to revert this
> change.

Well, the change jhb is suggesting won't reduce tc_lock contention for
sure, but it will prevent the turnstile lookup/locking if not
necessary.
Mainly, we need a spinlock held in order to protect mtx_lock bitfield
(as well as rw_lock) by setting waiters flags.
While in the past code this lock was only tc_lock, currently mtx_lock
is both protected by tc_lock and ts_lock, which it doesn't seem
strictly necessary.
if, for example, you lock tc_lock than you scan the chain and lock
ts_lock (basically what turnstile_trywait() does) and later you find
lock is released or that the state of lock changed and you cannot set
a waiters flag, you did the ts_locking and hash table scanning even if
it wasn't needed.

> I'm not sure why you're calling this O(n).  It's O(n) with the number of
> hash collisions, but it's a hash lookup.  We should consider using a real
> hash algorithm here if the distribution is bad.

Well, the hash lookup (worst case) has O(n) complexity.
This linear complexity cames from the O(1) * O(n) which is the access
to the list and the scanning of this one. So this is reduced to the
same complexity of a linked list.
However, some time ago I saw empirically that starting from a 'struct
thread' allocated address the highest level of entropy in the word
came after a right shift of 10 bits (for ia32) and 12 bits for amd64.
Currently, turnstile and sleepqueues only uses 8 bits of right
shifting.

Thanks,
Attilio


-- 
Peace can only be achieved by understanding - A. Einstein



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?3bbf2fe10709252359h7812ff7bsce9d393f80b65c06>