Re: [PATCH 5/6] futex: handle timeout inside adaptive lock spin

From: Darren Hart <dvhltc@us.ibm.com>
To: Gregory Haskins <ghaskins@novell.com>
Cc: Thomas Gleixner <tglx@linutronix.de>, Ingo Molnar <mingo@elte.hu>,
	Eric Dumazet <eric.dumazet@gmail.com>,
	Steven Rostedt <rostedt@goodmis.org>,
	Peter Zijlstra <peterz@infradead.org>,
	Peter Morreale <PMorreale@novell.com>,
	Sven Dietrich <SDietrich@novell.com>,
	Chris Mason <chris.mason@oracle.com>, Avi Kivity <avi@redhat.com>,
	Rik van Riel <riel@redhat.com>,
	Chris Wright <chrisw@sous-sol.org>,
	John Cooper <john.cooper@third-harmonic.com>,
	linux-kernel@vger.kernel.org
Subject: Re: [PATCH 5/6] futex: handle timeout inside adaptive lock	 spin
Date: Wed, 07 Apr 2010 16:15:05 -0700	[thread overview]
Message-ID: <4BBD11F9.1050705@us.ibm.com> (raw)
In-Reply-To: <4BBC9A3B0200005A00063E34@soto.provo.novell.com>

Gregory Haskins wrote:
 >>>> On 4/7/2010 at 01:31 PM, in message <4BBCC174.7020409@us.ibm.com>, 
Darren Hart
 > <dvhltc@us.ibm.com> wrote:
 >> Thomas Gleixner wrote:
 >>> On Mon, 5 Apr 2010, Darren Hart wrote:
 >>>
 >>>> Signed-off-by: Darren Hart <dvhltc@us.ibm.com>
 >
 >>>> +		if (timeout) {
 >>>> +			now = ktime_get();
 >>>   Hmm. Calling that in every iteration might hurt especially on non
 >>>   TSC systems, but well...
 >> I haven't come across a better alternative since arming the timer 
before
 >> setting TASK_INTERRUPTIBLE isn't appropriate.
 >
 > Hey Darren,
 >
 > I remember we tried something similar in early versions of the
 > adaptive locks and this was definitely bad. :(
 >
 > It ended up putting so much contention on the xtime_lock (IIRC) that
 > it resulted in adaptive locks hurting overall performance verses not
 > using adaptive at all.  Alternative mechanisms employed a hybrid where
 > the inner loops used a pseudo calibrated counter loop, and the outer
 > loop checks periodically against a real clock.  It all plays into "you
 > are burning CPU cycles anyway, so might as well put them to use"
 > theory.  Hacky, yes, but it did relieve the pressure on the time
 > subsystem locks and freed up a _lot_ of performance.  Without this,
 > the concept of timeouts+adaptive was unworkable.  I think Steven
 > ultimately rejected the timeout related patches outright when he
 > merged adaptive to -rt, but I think Sven pulled them into SLERT if you
 > would like a potential code reference to a working solution.
 >

Hi Greg,

Thanks for the info! I haven't tested with timeouts yet as I'm still 
struggling to get decent performance out of just plain old adaptive 
right now. I'll keep that in mind when I get to that, and yeah, if the 
plan is to burn CPU cycles, might as well do something constructive :-)

I do feel the timeouts are a necessary feature. Interruptibility may be 
as well, but I'm going to ignore it for the time being...

--
Darren