835695 - Use jar reordering with core apps packages

Status: RESOLVED FIXED

(Firefox OS Graveyard :: General, defect)

Description

[[:fabrice] Fabrice Desré]

This is an experiment to see if reordering application packages helps with startup time.

Comment 1

[[:fabrice] Fabrice Desré]

Attachment: gecko patch v1

Comment 2

[[:fabrice] Fabrice Desré]

Attachment: gaia patch v1

To set up everything properly, apply the patches, and flash a VARIANT=user build.

Then run these steps:
- in gaia, make start-reorder
- launch all the apps on the phone that you want to preload
- in gaia, make stop-reorder
- in gaia, PRODUCTION=1 make production

Comment 3

[[:fabrice] Fabrice Desré]

Chris, do you mind running your timing tests with these?

Comment 4

[Chris Jones [:cjones] inactive; ni?/f?/r? if you need me]

mwu, can you help measure this.

Comment 5

[Michael Wu [:mwu]]

I started with trying to get a measure of how much IO contributes to the startup time, but startup time seems to massively vary regardless of whether we have hot or cold caches. I'll try measuring with reordered jars but I doubt we can see much improvement if a hot cache doesn't necessary give us better startup..

Comment 6

[Chris Jones [:cjones] inactive; ni?/f?/r? if you need me]

What steps are you using to test?  What app are you testing?

Comment 7

[Michael Wu [:mwu]]

I updated gecko and changed my testing procedure a bit, and the numbers are a lot less noisy now. Email had a difference of ~130ms between hot and cold startup, and clock had a difference of ~200ms. However, even after reordering, the startup times did not seem to change much.

Comment 8

[[:fabrice] Fabrice Desré]

So I revisited that technique and results are pretty different, probably because gaia apps moved from loading a few big files to loading a lot of small ones and it's more important to reorder.

On a tarako, cold launch of the settings apps drops from ~2000ms to ~1720ms - a huge gain for "just" some build time machinery.

Yuren, let's work on that next week together!

Comment 9

[[:fabrice] Fabrice Desré]

Attachment: Gaia PR for v1.3t

Comment 10

[[:fabrice] Fabrice Desré]

Attachment: gonk-misc patch

Michael, this change to b2g.sh let us restart in jar reordering logging mode by setting a temporary property. Looks cleaner to me than to pull/push b2g.sh to add the export on the fly.

Comment 11

[[:fabrice] Fabrice Desré]

Attachment: reorder-b2g-sh.patch

Updated according to irc discussion.

Comment 12

[Michael Wu [:mwu]]

Comment on attachment 8373756 [details] [diff] [review]
reorder-b2g-sh.patch

Review of attachment 8373756 [details] [diff] [review]:
-----------------------------------------------------------------

r=me on the jar part

::: /home/fabrice/dev/tarako/gonk-misc/b2g.sh
@@ -4,4 @@
>  mkdir -p $TMPDIR
>  chmod 1777 $TMPDIR
>  ulimit -n 8192
> -export MOZ_CRASHREPORTER=1

?

Comment 13

[[:fabrice] Fabrice Desré]

https://github.com/mozilla-b2g/gonk-misc/commit/78b908b493bfe0b477e3d4f6edec8c46a2c0d096

Comment 14

[Yuren [:yurenju] (AFK ~ Oct. 12) (volunteer mode)]

Comment on attachment 8373219 [details] [review]
Gaia PR for v1.3t

r=yurenju and please add a space around | for this line:
> APPS=`cut -f 1 -d ' ' jarloader.log |sort -u|grep gaiamobile| cut -c 8-`

for 1.3t we can land this patch but for master we should rewrite optimizejar.py in javascript. 

do you think we need a follow up bug to rewrite in javascript for master or just trace it in this bug?

Comment 15

[[:fabrice] Fabrice Desré]

Thanks Yuren! We should definitely do the js rewrite in another bug.

Comment 16

[[:fabrice] Fabrice Desré]

1.3t: https://github.com/mozilla-b2g/gaia/commit/14cf881c8559082308b8f68b1353499fdc87bbe5

Comment 17

[[:fabrice] Fabrice Desré]

log files on 1.3t:
https://github.com/mozilla-b2g/gaia/commit/56bd97aef3e453a008ee2c1908dce16e7afd127e