RE: [RFC 0/6] optimize ctx switch with rb-tree

From: "Liang, Kan" <kan.liang@intel.com>
To: David Carrillo-Cisneros <davidcc@google.com>,
	"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>
Cc: "x86@kernel.org" <x86@kernel.org>, Ingo Molnar <mingo@redhat.com>,
	"Thomas Gleixner" <tglx@linutronix.de>,
	Andi Kleen <ak@linux.intel.com>,
	"Peter Zijlstra" <peterz@infradead.org>,
	Borislav Petkov <bp@suse.de>,
	"Srinivas Pandruvada" <srinivas.pandruvada@linux.intel.com>,
	Dave Hansen <dave.hansen@linux.intel.com>,
	Vikas Shivappa <vikas.shivappa@linux.intel.com>,
	Mark Rutland <mark.rutland@arm.com>,
	Arnaldo Carvalho de Melo <acme@kernel.org>,
	Vince Weaver <vince@deater.net>, Paul Turner <pjt@google.com>,
	Stephane Eranian <eranian@google.com>,
	"Budankov, Alexey" <alexey.budankov@intel.com>
Subject: RE: [RFC 0/6] optimize ctx switch with rb-tree
Date: Tue, 25 Apr 2017 17:27:55 +0000	[thread overview]
Message-ID: <37D7C6CF3E00A74B8858931C1DB2F077536E6010@SHSMSX103.ccr.corp.intel.com> (raw)
In-Reply-To: <20170110102502.106187-1-davidcc@google.com>

Hi David,

Is there any update about the patch series?

We recently encountered another performance issue on KNL. I think the RB-tree solution also has benefits for it.

Thanks,
Kan

> Subject: [RFC 0/6] optimize ctx switch with rb-tree
> 
> Following the discussion in:
> https://patchwork.kernel.org/patch/9420035/
> 
> This is is an early version of a series of perf context switches optimizations.
> 
> The main idea is to create and maintain a list of inactive events sorted by
> timestamp, and a rb-tree index to index it. The rb-tree's key are
> {cpu,flexible,stamp} for task contexts and {cgroup,flexible,stamp} for CPU
> contexts.
> 
> The rb-tree provides functions to find intervals in the inactive event list so
> that ctx_sched_in only has to visit the events that can be potentially be
> scheduled (i.e. avoid iterations over events bound to CPUs or cgroups that
> are not current).
> 
> Since the inactive list is sort by timestamp, rotation can be done by simply
> scheduling out and in the events. This implies that each timer interrupt, the
> events will rotate by q events (where q is the number of hardware counters).
> This changes the current behavior of rotation.
> Feedback welcome!
> 
> I haven't profiled the new approach. I am only assuming it will be superior
> when the number of per-cpu or distict cgroup events is large.
> 
> The last patch shows how perf_iterate_ctx can use the new rb-tree index to
> reduce the number of visited events. I haven't looked carefully if locking and
> other things are correct.
> 
> If this changes are in the right direction. A next version could remove some
> existing code, specifically the lists ctx->pinned_groups and
> ctx->flexible_groups could be removed. Also, event_filter_match could be
> simplified when called on events groups filtered using the rb-tree, since both
> perform similar checks.
> 
> David Carrillo-Cisneros (6):
>   perf/core: create active and inactive event groups
>   perf/core: add a rb-tree index to inactive_groups
>   perf/core: use rb-tree to sched in event groups
>   perf/core: avoid rb-tree traversal when no inactive events
>   perf/core: rotation no longer neccesary. Behavior has changed. Beware
>   perf/core: use rb-tree index to optimize filtered  perf_iterate_ctx
> 
>  include/linux/perf_event.h |  13 ++
>  kernel/events/core.c       | 466
> +++++++++++++++++++++++++++++++++++++++------
>  2 files changed, 426 insertions(+), 53 deletions(-)
> 
> --
> 2.11.0.390.gc69c2f50cf-goog