Boot prefetch optimization for Chrome Frame (experimental)...

Boot prefetch optimization for Chrome Frame (experimental)

If chrome is warmed up during a single reboot, it gets paged
in for subsequent reboots and the cold startup times essentially
look like warm times thereafter! The 'warm up' is done by
setting up a 'RunOnce' key during DLLRegisterServer of 
npchrome_frame.dll.

This works because chrome prefetch becomes part of boot
prefetch file ntosboot-b00dfaad.pf and paged in on subsequent
reboots. As long as the sytem does not undergo significant 
memory pressure those pages remain in memory and we get pretty
amazing startup times, down to about 300 ms from 1200 ms

The downside is:
- Whether chrome frame is used or not, there's a read penalty
  (1200-300 =) 900 ms for every boot.
- Heavy system memory usage after reboot will nullify the benefits
  but the user will still pay the cost.
- Overall the time saved will always be less than total time spent
  paging in chrome
- We are not sure when the chrome 'warm up' will age out from the
  boot prefetch file.

The idea here is to try this out on chrome frame dev channel
and see if it produces a significant drift in startup numbers.



Committed: http://src.chromium.org/viewvc/chrome?view=rev&revision=52489

[slightlyoff, 2010-07-12 23:49:36]

Thanks for writing up our research so clearly.

http://codereview.chromium.org/2936010/diff/1/2
File chrome_frame/chrome_tab.cc (right):

http://codereview.chromium.org/2936010/diff/1/2#newcode306
chrome_frame/chrome_tab.cc:306: if
(!GoogleUpdateSettings::GetChromeChannel(true, &channel_name) ||
This check is unnecessary. If it doesn't pan out on Dev, we can just remove
this.

http://codereview.chromium.org/2936010/diff/1/2#newcode312
chrome_frame/chrome_tab.cc:312: if (run_once.Open(HKEY_LOCAL_MACHINE, kRunOnce,
KEY_READ | KEY_WRITE)) {
why not HKCU?

[tommi, 2010-07-13 15:04:15]

lgtm with alex' comments

http://codereview.chromium.org/2936010/diff/1/2
File chrome_frame/chrome_tab.cc (right):

http://codereview.chromium.org/2936010/diff/1/2#newcode301
chrome_frame/chrome_tab.cc:301: HRESULT SetupRunOnce() {
Can you expand this comment to include the gist of the change list description? 
Otherwise it's not clear how this works from the code.

http://codereview.chromium.org/2936010/diff/1/2#newcode315
chrome_frame/chrome_tab.cc:315: switches::kAutomationClientChannelID, _T("0"));
L instead of _T()?

[amit, 2010-07-14 22:06:32]

http://codereview.chromium.org/2936010/diff/1/2
File chrome_frame/chrome_tab.cc (right):

http://codereview.chromium.org/2936010/diff/1/2#newcode306
chrome_frame/chrome_tab.cc:306: if
(!GoogleUpdateSettings::GetChromeChannel(true, &channel_name) ||
On 2010/07/12 23:49:36, slightlyoff wrote:
> This check is unnecessary. If it doesn't pan out on Dev, we can just remove
> this.
Belt and suspenders :) 
Branch date is near and we might need to run this for a few weeks to get a
meaningful feedback. I don't want this to be enabled for anything else but dev.

http://codereview.chromium.org/2936010/diff/1/2#newcode312
chrome_frame/chrome_tab.cc:312: if (run_once.Open(HKEY_LOCAL_MACHINE, kRunOnce,
KEY_READ | KEY_WRITE)) {
On 2010/07/12 23:49:36, slightlyoff wrote:
> why not HKCU?

Done.

http://codereview.chromium.org/2936010/diff/1/2#newcode315
chrome_frame/chrome_tab.cc:315: switches::kAutomationClientChannelID, _T("0"));
On 2010/07/13 15:04:16, tommi wrote:
> L instead of _T()?

Done.

[slightlyoff, 2010-07-15 01:04:40]

LGTM

On 2010/07/14 22:06:32, amit wrote:
> http://codereview.chromium.org/2936010/diff/1/2
> File chrome_frame/chrome_tab.cc (right):
> 
> http://codereview.chromium.org/2936010/diff/1/2#newcode306
> chrome_frame/chrome_tab.cc:306: if
> (!GoogleUpdateSettings::GetChromeChannel(true, &channel_name) ||
> On 2010/07/12 23:49:36, slightlyoff wrote:
> > This check is unnecessary. If it doesn't pan out on Dev, we can just remove
> > this.
> Belt and suspenders :) 
> Branch date is near and we might need to run this for a few weeks to get a
> meaningful feedback. I don't want this to be enabled for anything else but
dev.
> 
> http://codereview.chromium.org/2936010/diff/1/2#newcode312
> chrome_frame/chrome_tab.cc:312: if (run_once.Open(HKEY_LOCAL_MACHINE,
kRunOnce,
> KEY_READ | KEY_WRITE)) {
> On 2010/07/12 23:49:36, slightlyoff wrote:
> > why not HKCU?
> 
> Done.
> 
> http://codereview.chromium.org/2936010/diff/1/2#newcode315
> chrome_frame/chrome_tab.cc:315: switches::kAutomationClientChannelID,
_T("0"));
> On 2010/07/13 15:04:16, tommi wrote:
> > L instead of _T()?
> 
> Done.