[mojo-edk] Cache the context thread-local in the RequestContext.

[mojo-edk] Cache the context thread-local in the RequestContext.

On MojoBindingsPerftest.InProcessPingPong, RequestContext's constructor
and destructor combined consume ~22% of CPU time. The reason is that
accessing a LazyInstance compiles down to a number of atomic
operations, which are unnecessarily done twice in each of the
constructor and destructor. Since the LazyInstance is leaky, it is safe
to read the pointer once and cache it every time we need to access the
thread-local. This change results in ~20% improvement in
MojoBindingsPerftest.InProcessPingPong performance.

The CPU usage is smaller but still noticable in MojoE2EPerftest with
small messages. With a batch size of 1, this consumes ~4.5% of CPU time,
rising to ~9% with a batch size of 10.

BUG=590495
Committed: https://crrev.com/c8e9c39cfb8bef5e49c2e2b1437267eb165e2dae
Cr-Commit-Position: refs/heads/master@{#379513}

[Anand Mistry (off Chromium), 2016-03-07 04:21:40]

amistry@chromium.org changed reviewers:
+ rockot@chromium.org

[Ken Rockot(use gerrit already), 2016-03-07 04:25:00]

https://codereview.chromium.org/1767003002/diff/1/mojo/edk/system/request_con...
File mojo/edk/system/request_context.h (right):

https://codereview.chromium.org/1767003002/diff/1/mojo/edk/system/request_con...
mojo/edk/system/request_context.h:83: base::ThreadLocalPointer<RequestContext>*
tls_context_;
Hmm, couldn't this just be a raw pointer?

[Ken Rockot(use gerrit already), 2016-03-07 04:25:22]

On 2016/03/07 at 04:25:00, Ken Rockot wrote:
>
https://codereview.chromium.org/1767003002/diff/1/mojo/edk/system/request_con...
> File mojo/edk/system/request_context.h (right):
> 
>
https://codereview.chromium.org/1767003002/diff/1/mojo/edk/system/request_con...
> mojo/edk/system/request_context.h:83:
base::ThreadLocalPointer<RequestContext>* tls_context_;
> Hmm, couldn't this just be a raw pointer?

Oh wait, nevermind. It is a raw pointer TO a TLP. Missed that. LGTM

[Anand Mistry (off Chromium), 2016-03-07 05:04:31]

On 2016/03/07 04:25:22, Ken Rockot wrote:
> On 2016/03/07 at 04:25:00, Ken Rockot wrote:
> >
>
https://codereview.chromium.org/1767003002/diff/1/mojo/edk/system/request_con...
> > File mojo/edk/system/request_context.h (right):
> > 
> >
>
https://codereview.chromium.org/1767003002/diff/1/mojo/edk/system/request_con...
> > mojo/edk/system/request_context.h:83:
> base::ThreadLocalPointer<RequestContext>* tls_context_;
> > Hmm, couldn't this just be a raw pointer?
> 
> Oh wait, nevermind. It is a raw pointer TO a TLP. Missed that. LGTM

There's something else going on, so I've commented on crbug.com/423074. Each
LazyInstance access should be a single mfence in the Acquire_Load on the fast
path, but for some reason, there are 4. So each RequestContext on the stack
costs 16 mfence's. This change drops it to 4, and theoretically 1 if
Acquire_Load and NoBarrier_Load are fixed.

[Anand Mistry (off Chromium), 2016-03-07 05:04:57]

The CQ bit was checked by amistry@chromium.org

[commit-bot: I haz the power, 2016-03-07 05:05:03]

CQ is trying da patch. Follow status at
 https://chromium-cq-status.appspot.com/patch-status/1767003002/1
View timeline at
 https://chromium-cq-status.appspot.com/patch-timeline/1767003002/1

[commit-bot: I haz the power, 2016-03-07 05:27:32]

Committed patchset #1 (id:1)

[commit-bot: I haz the power, 2016-03-07 05:28:53]

Description was changed from

==========
[mojo-edk] Cache the context thread-local in the RequestContext.

On MojoBindingsPerftest.InProcessPingPong, RequestContext's constructor
and destructor combined consume ~22% of CPU time. The reason is that
accessing a LazyInstance compiles down to a number of atomic
operations, which are unnecessarily done twice in each of the
constructor and destructor. Since the LazyInstance is leaky, it is safe
to read the pointer once and cache it every time we need to access the
thread-local. This change results in ~20% improvement in
MojoBindingsPerftest.InProcessPingPong performance.

The CPU usage is smaller but still noticable in MojoE2EPerftest with
small messages. With a batch size of 1, this consumes ~4.5% of CPU time,
rising to ~9% with a batch size of 10.

BUG=590495
==========

to

==========
[mojo-edk] Cache the context thread-local in the RequestContext.

On MojoBindingsPerftest.InProcessPingPong, RequestContext's constructor
and destructor combined consume ~22% of CPU time. The reason is that
accessing a LazyInstance compiles down to a number of atomic
operations, which are unnecessarily done twice in each of the
constructor and destructor. Since the LazyInstance is leaky, it is safe
to read the pointer once and cache it every time we need to access the
thread-local. This change results in ~20% improvement in
MojoBindingsPerftest.InProcessPingPong performance.

The CPU usage is smaller but still noticable in MojoE2EPerftest with
small messages. With a batch size of 1, this consumes ~4.5% of CPU time,
rising to ~9% with a batch size of 10.

BUG=590495

Committed: https://crrev.com/c8e9c39cfb8bef5e49c2e2b1437267eb165e2dae
Cr-Commit-Position: refs/heads/master@{#379513}
==========

[commit-bot: I haz the power, 2016-03-07 05:28:54]

Patchset 1 (id:??) landed as
https://crrev.com/c8e9c39cfb8bef5e49c2e2b1437267eb165e2dae
Cr-Commit-Position: refs/heads/master@{#379513}