Re: [RFC patch 0/5] futex: Allow lockless empty check of hashbucket plist in futex_wake()

From: Davidlohr Bueso <davidlohr@hp.com>
To: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>,
	LKML <linux-kernel@vger.kernel.org>,
	Jason Low <jason.low2@hp.com>, Ingo Molnar <mingo@kernel.org>,
	Darren Hart <dvhart@linux.intel.com>,
	Mike Galbraith <efault@gmx.de>, Jeff Mahoney <jeffm@suse.com>,
	Linus Torvalds <torvalds@linux-foundation.org>,
	Scott Norton <scott.norton@hp.com>, Tom Vaden <tom.vaden@hp.com>,
	Aswin Chandramouleeswaran <aswin@hp.com>,
	Waiman Long <Waiman.Long@hp.com>,
	"Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
Subject: Re: [RFC patch 0/5] futex: Allow lockless empty check of hashbucket plist in futex_wake()
Date: Sat, 30 Nov 2013 20:37:33 -0800	[thread overview]
Message-ID: <1385872653.17749.13.camel@buesod1.americas.hpqcorp.net> (raw)
In-Reply-To: <20131128115946.GD13532@twins.programming.kicks-ass.net>

On Thu, 2013-11-28 at 12:59 +0100, Peter Zijlstra wrote:
> On Wed, Nov 27, 2013 at 11:44:38PM -0800, Davidlohr Bueso wrote:
> > How about both enlarging the table _and_ aligning the buckets? As you
> > know, increasing the size of the table also benefits (particularly in
> > larger systems) in having more spinlocks. So we reduce the amount of
> > collisions and alleviate contention on the hb->lock. Btw, do you have
> > any particular concerns about the larger hash table patch? 
> 
> My only concern was the amount of #ifdef.
> 
> Wouldn't something like the below also work?

Below are the results for a workload that stresses the uaddr hashing for
large amounts of futexes (just make waits fail the uval check, so no
list handing overhead) on an 80 core, 1Tb NUMA system.

+---------+--------------------+------------------------+-----------------------+-------------------------------+
| threads | baseline (ops/sec) | aligned-only (ops/sec) | large table (ops/sec) | large table+aligned (ops/sec) |
+---------+--------------------+------------------------+-----------------------+-------------------------------+
|     512 |              32426 | 50531  (+55.8%)        | 255274  (+687.2%)     | 292553  (+802.2%)             |
|     256 |              65360 | 99588  (+52.3%)        | 443563  (+578.6%)     | 508088  (+677.3%)             |
|     128 |             125635 | 200075 (+59.2%)        | 742613  (+491.1%)     | 835452  (+564.9%)             |
|      80 |             193559 | 323425 (+67.1%)        | 1028147 (+431.1%)     | 1130304 (+483.9%)             |
|      64 |             247667 | 443740 (+79.1%)        | 997300  (+302.6%)     | 1145494 (+362.5%)             |
|      32 |             628412 | 721401 (+14.7%)        | 965996  (+53.7%)      | 1122115 (+78.5%)              |
+---------+--------------------+------------------------+-----------------------+-------------------------------+

Baseline of course sucks compared to any other performance boost, and we
get the best throughput when applying both optimizations, no surprise.
We do particularly well for more than 32 threads, and the 'aligned-only'
column nicely exemplifies the benefits of SMP aligning the buckets
without considering the reduction in collisions.

Thanks,
Davidlohr