get() / set() perf optimizations #17166

kanongil · 2018-10-31T17:31:26Z

I noticed that the ember-performance Chrome scores of get() and set() dropped dramatically when ember-metal-es5-getters was enabled, so I had a look at it.

I have managed to mostly bring it back for get(), and improve it for set() using the included patches. Both are roughly 2x the canary performance. Additionally, it seems to have improved the run() benchmark as well.

All the patches should work individually, in case you only think some of them are suitable.

Chrome

User Agent: Mozilla/5.0 (Macintosh; Intel Mac OS X 10_14_1) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/71.0.3578.20 Safari/537.36

.------------------------------------------------------------------------.
|                   Ember Performance Suite - Results                    |
|------------------------------------------------------------------------|
|         Name         |       Speed        | Error  | Samples |  Mean   |
|----------------------|--------------------|--------|---------|---------|
|  -- Ember 3.0.0 --   |                    |        |         |         |
| Ember.get            | 1,626,872.70 / sec | ∓1.08% | 64      | 0.00 ms |
| Ember.run            | 670,554.44 / sec   | ∓2.42% | 65      | 0.00 ms |
| Ember.set            | 606,311.51 / sec   | ∓1.12% | 64      | 0.00 ms |
|  -- Ember 3.4.5 --   |                    |        |         |         |
| Ember.get            | 770,484.17 / sec   | ∓1.65% | 61      | 0.00 ms |
| Ember.run            | 710,395.41 / sec   | ∓6.42% | 58      | 0.00 ms |
| Ember.set            | 408,123.26 / sec   | ∓0.47% | 65      | 0.00 ms |
|  -- Ember canary --  |                    |        |         |         |
| Ember.get            | 858,315.30 / sec   | ∓0.96% | 61      | 0.00 ms |
| Ember.run            | 739,482.59 / sec   | ∓0.34% | 66      | 0.00 ms |
| Ember.set            | 453,404.65 / sec   | ∓0.61% | 65      | 0.00 ms |
|  -- Ember dev --     |                    |        |         |         |
| Ember.get            | 1,487,276.86 / sec | ∓1.04% | 63      | 0.00 ms |
| Ember.run            | 818,811.25 / sec   | ∓0.53% | 66      | 0.00 ms |
| Ember.set            | 996,207.06 / sec   | ∓1.33% | 61      | 0.00 ms |
'------------------------------------------------------------------------'

Safari

User Agent: Mozilla/5.0 (Macintosh; Intel Mac OS X 10_14_1) AppleWebKit/605.1.15 (KHTML, like Gecko) Version/12.0.1 Safari/605.1.15

.------------------------------------------------------------------------.
|                   Ember Performance Suite - Results                    |
|------------------------------------------------------------------------|
|         Name         |       Speed        | Error  | Samples |  Mean   |
|----------------------|--------------------|--------|---------|---------|
|  -- Ember 3.0.0 --   |                    |        |         |         |
| Ember.get            | 1,024,541.17 / sec | ∓0.93% | 62      | 0.00 ms |
| Ember.run            | 216,263.15 / sec   | ∓3.47% | 37      | 0.00 ms |
| Ember.set            | 711,296.65 / sec   | ∓0.99% | 64      | 0.00 ms |
|  -- Ember 3.4.5 --   |                    |        |         |         |
| Ember.get            | 660,504.61 / sec   | ∓0.83% | 62      | 0.00 ms |
| Ember.run            | 227,034.91 / sec   | ∓3.44% | 36      | 0.00 ms |
| Ember.set            | 393,704.49 / sec   | ∓0.93% | 63      | 0.00 ms |
|  -- Ember canary --  |                    |        |         |         |
| Ember.get            | 888,268.20 / sec   | ∓0.96% | 61      | 0.00 ms |
| Ember.run            | 236,605.91 / sec   | ∓3.75% | 36      | 0.00 ms |
| Ember.set            | 490,658.82 / sec   | ∓0.80% | 63      | 0.00 ms |
|  -- Ember dev --     |                    |        |         |         |
| Ember.get            | 952,592.90 / sec   | ∓0.96% | 63      | 0.00 ms |
| Ember.run            | 256,341.28 / sec   | ∓3.60% | 37      | 0.00 ms |
| Ember.set            | 690,086.43 / sec   | ∓1.09% | 62      | 0.00 ms |
'------------------------------------------------------------------------'

kanongil · 2018-10-31T17:33:14Z

packages/@ember/-internals/meta/lib/meta.ts

@@ -480,16 +494,15 @@ export class Meta {
    while (pointer !== null) {
      let map = pointer._descriptors;
      if (map !== undefined) {
-        for (let key in map) {
+        map.forEach((value, key) => {


Note: I use Map.forEach() for the iteration, since it is the only method that works with IE11.

rwjblue · 2018-10-31T18:14:47Z

These changes all seem pretty good to me, thanks for working on this!

rwjblue · 2018-10-31T18:14:59Z

@krisselden - Mind taking a look?

kanongil · 2018-10-31T22:30:16Z

It might also be possible to remove the descriptorFor() call in get(), since the value can simply be obtained by accessing the attribute (calling any getter). I skipped it for now – mostly because I'm not sure if there are any implications I'm missing from doing this.

rwjblue · 2018-11-01T01:17:35Z

Looks like some linting failures, mind taking a look?

bekzod · 2018-11-01T07:49:20Z

packages/@ember/-internals/metal/lib/property_events.ts


-  if (hasMeta && (meta.isInitializing() || meta.isPrototypeMeta(obj))) {
+  if (meta !== null && (meta.isInitializing() || meta.isPrototypeMeta(obj))) {


why not check once and reference it like before let hasMeta = meta !== null ?

Because the ts linter won't recognize that meta is non-null, and the subsequent statements would need a "!":

if (hasMeta && (meta!.isInitializing() || meta!.isPrototypeMeta(obj))) {

bekzod · 2018-11-01T07:55:14Z

packages/@ember/-internals/metal/lib/property_get.ts

  let obj: any = root;
-  let parts = path.split('.');
+  let parts = typeof path === 'string' ? path.split('.') : path;


I think it will be even more performant if _getPath only accept array of path, https://github.com/emberjs/ember.js/pull/17166/files#diff-45589baa0b0954cefba717cf63c270f9R100 and here you can put isObjectLike ? _getPath(obj, keyName.split('.')) : undefined

Yes, I thought about that, and you might be right. I discarded it as a micro optimization I didn't want to do in this pass. It is not likely to make any measurable difference.

bekzod · 2018-11-01T07:56:56Z

I had similar approach in https://github.com/emberjs/ember.js/pull/17057/files but gave up since it was linting nightmare :)

bekzod · 2018-11-01T08:04:29Z

this makes #17058 redundant since it is addressed here already

kanongil · 2018-11-01T09:30:32Z

@bekzod I actually found the typescript very helpful doing the patch. I doubt I would have dared to do it without.

bekzod · 2018-11-01T09:56:31Z

of course typescript is very helpful for such refactors, but when I did #17057 it required a lot of ! signs had similar issues #17166 (comment)

kanongil · 2018-11-11T14:46:59Z

@rwjblue I fixed the linting issue.

rwjblue · 2018-12-07T02:07:47Z

@kanongil we chatted about this a bit in discord a week or two ago, you were working on getting ember-macro-benchmark running with this change to ensure there is not a regression outside of microbenchmarks. Have you had a chance to do that yet?

kanongil · 2018-12-07T13:42:27Z

Yes, sorry about the delay. It took quite a bit of effort to get ember-macro-benchmark to work properly.

Initially, it wouldn't patch the ember source into the project, but I think I got it solved. I managed to get it to work with emberaddons.com and this config:

{
  "har": "archives/www.emberaddons.com.har",
  "runCount": 100,
  "cpuThrottleRate": 4,
  "servers": [{
    "name": "master-e2352e6",
    "port": 8880,
    "ember": "embers/ember-master-e2352e6.min.js"
  }, {
    "name": "get-set-perf-d6e5472",
    "port": 8881,
    "ember": "embers/ember-get-set-perf-d6e5472.min.js"
  }]
}

Result: results.pdf

rwjblue · 2018-12-07T15:14:24Z

OK, great, so the results definitely don't show a regression (suggests that the change is statistically insignificant for emberaddons.com).

IMHO, this code is strictly better than what we had before which means that as long as we do our due diligence to confirm it does not introduce a regression (which it seems that we have) we should land...

rwjblue · 2018-12-07T15:15:46Z

@krisselden - Do you to do one more pass at review before landing?

rwjblue · 2018-12-07T22:23:46Z

Looks like a recent PR merged made this have conflicts, mind fixing that up?

This avoids redundant peeking when passing empty meta to certain methods

rwjblue · 2019-01-02T20:24:03Z

Just rebased to fix conflicts, will merge once CI is done...

rwjblue · 2019-01-02T22:17:04Z

@kanongil - Thank you for pushing this forward!

kanongil commented Oct 31, 2018

View reviewed changes

rwjblue requested a review from krisselden October 31, 2018 18:13

bekzod reviewed Nov 1, 2018

View reviewed changes

rwjblue approved these changes Dec 7, 2018

View reviewed changes

kanongil added 5 commits January 2, 2019 15:14

Use null to distinguish between missing or empty meta param

1eb257e

This avoids redundant peeking when passing empty meta to certain methods

Use Map() for descriptor store

5cf6543

Only call split() once and avoid a join() for setPath()

a51c182

Resolve path names correctly

cbbc4c4

Linting fixes

89acaf9

rwjblue force-pushed the get-set-perf branch from d6e5472 to 89acaf9 Compare January 2, 2019 20:23

rwjblue merged commit c2623b9 into emberjs:master Jan 2, 2019

bekzod mentioned this pull request Jan 31, 2019

chore: fail test for each-in not accessing props properly whose keys contain a period #17535

Closed

pzuraq mentioned this pull request Sep 11, 2019

Ember.get/Ember.getWithDefault don't work as expected when the key name has "." in it since version 3.8 #18377

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

get() / set() perf optimizations #17166

get() / set() perf optimizations #17166

kanongil commented Oct 31, 2018

kanongil Oct 31, 2018

rwjblue commented Oct 31, 2018

rwjblue commented Oct 31, 2018

kanongil commented Oct 31, 2018

rwjblue commented Nov 1, 2018

bekzod Nov 1, 2018

kanongil Nov 1, 2018

bekzod Nov 1, 2018

kanongil Nov 1, 2018

bekzod commented Nov 1, 2018

bekzod commented Nov 1, 2018

kanongil commented Nov 1, 2018

bekzod commented Nov 1, 2018

kanongil commented Nov 11, 2018

rwjblue commented Dec 7, 2018

kanongil commented Dec 7, 2018

rwjblue commented Dec 7, 2018

rwjblue commented Dec 7, 2018

rwjblue commented Dec 7, 2018

rwjblue commented Jan 2, 2019

rwjblue commented Jan 2, 2019


		if (hasMeta && (meta.isInitializing() \|\| meta.isPrototypeMeta(obj))) {
		if (meta !== null && (meta.isInitializing() \|\| meta.isPrototypeMeta(obj))) {

get() / set() perf optimizations #17166

get() / set() perf optimizations #17166

Conversation

kanongil commented Oct 31, 2018

kanongil Oct 31, 2018

Choose a reason for hiding this comment

rwjblue commented Oct 31, 2018

rwjblue commented Oct 31, 2018

kanongil commented Oct 31, 2018

rwjblue commented Nov 1, 2018

bekzod Nov 1, 2018

Choose a reason for hiding this comment

kanongil Nov 1, 2018

Choose a reason for hiding this comment

bekzod Nov 1, 2018

Choose a reason for hiding this comment

kanongil Nov 1, 2018

Choose a reason for hiding this comment

bekzod commented Nov 1, 2018

bekzod commented Nov 1, 2018

kanongil commented Nov 1, 2018

bekzod commented Nov 1, 2018

kanongil commented Nov 11, 2018

rwjblue commented Dec 7, 2018

kanongil commented Dec 7, 2018

rwjblue commented Dec 7, 2018

rwjblue commented Dec 7, 2018

rwjblue commented Dec 7, 2018

rwjblue commented Jan 2, 2019

rwjblue commented Jan 2, 2019