-
Notifications
You must be signed in to change notification settings - Fork 30k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
process: improve cwd performance #27224
Conversation
This comment has been minimized.
This comment has been minimized.
I think the native addons situation does make this semver-major, at least, although it’s unlikely that many native addons change the current working directory. I’m not sure what the most expensive part of caling
With the approach in this PR, you’d have to notify workers when the cwd changes – for low overhead, you could e.g. use |
@joyeecheung I tried to add a I circumvented this issue by only improving the performance for the main thread for now but it would actually be nice to improve workers as well. |
This comment has been minimized.
This comment has been minimized.
@BridgeAR I think what you need is communication between worker.js and main/worker_thread.js? Or you can expose a wrapper in process/execution.js. Why do you need to go through bootstrap.js? A SharedArrayBuffer containing something like cwd is not supposed to be part of the snapshot. |
@joyeecheung the // Worker thread
let cachedCwd = realMethods.cwd();
let lastCounter = -1;
function cwd() {
const currentCounter = Atomics.load(counter, 0);
if (currentCounter === lastCounter)
return cachedCwd;
lastCounter = currentCounter;
cachedCwd = realMethods.cwd();
return cachedCwd;
}
// Main thread only.
// This buffer would have to be shared and passed through to each worker.
const counter = new Uint64Array(new SharedArrayBuffer(8));
function chdir(cwd) {
Atomics.add(counter, 0, 1);
realMethods.chdir(cwd);
} I could theoretically override the two functions in case the first worker is started and to use a shared buffer at that point but AFAIK we do not want to override the functions anymore at that point, right? |
Local benchmark: // Before
console.time(); for (var i = 0; i < 1e7; i++) process.cwd(); console.timeEnd()
// default: 5596.328ms
// After
console.time(); for (var i = 0; i < 1e7; i++) process.cwd(); console.timeEnd()
// default: 477.811ms |
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
@BridgeAR You could delay the creation of |
um, wait, in case of workers, I suppose the SAB has to be sent together with the |
This comment has been minimized.
This comment has been minimized.
I just updated the PR to also work with workers. |
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
@nodejs/tsc PTAL. This needs one more review from the TSC. |
This caches the current working directory and only updates the variable if `process.chdir()` is called. PR-URL: nodejs#27224 Reviewed-By: John-David Dalton <john.david.dalton@gmail.com> Reviewed-By: Anna Henningsen <anna@addaleax.net> Reviewed-By: Michaël Zasso <targos@protonmail.com>
Landed in 1d022e8 🎉 |
This caches the current working directory and only updates the variable if `process.chdir()` is called. PR-URL: nodejs#27224 Reviewed-By: John-David Dalton <john.david.dalton@gmail.com> Reviewed-By: Anna Henningsen <anna@addaleax.net> Reviewed-By: Michaël Zasso <targos@protonmail.com>
This makes sure nodejs#27224 is possible to being backported in a semver-patch way.
This caches the current working directory and only updates the variable if `process.chdir()` is called. PR-URL: #27224 Reviewed-By: John-David Dalton <john.david.dalton@gmail.com> Reviewed-By: Anna Henningsen <anna@addaleax.net> Reviewed-By: Michaël Zasso <targos@protonmail.com> Backport-PR-URL: #27483
This makes sure #27224 is possible to being backported in a semver-patch way. PR-URL: #27483 Reviewed-By: Anna Henningsen <anna@addaleax.net> Reviewed-By: Michaël Zasso <targos@protonmail.com> Reviewed-By: Beth Griggs <Bethany.Griggs@uk.ibm.com> Reviewed-By: Joyee Cheung <joyeec9h3@gmail.com> Reviewed-By: Shelley Vohr <codebytere@gmail.com>
This commit improves the lookups of `ignore` and `pending` configs by reducing the amount of iterations required over those config properties. tl;dr: - cached the config lookups for `pending` and `ignore` to reduce the amount of times we have to re-iterate the `pending` or `ignore` arrays provided by the config. - cache string lookups for normalized moduleId paths since This was done in three steps: First Step The normalized `moduleId` was cached. Running ember-template-lint on a big codebase (668 template files) took 12.94 seconds to run. I used this command in the app's codebase directory to profile: ``` node --prof ../ember-template-lint/bin/ember-template-lint.js . ``` Next, I generated a graph with the [flamebearer][flamebearer] npm package: ``` node --prof-process --preprocess -j isolate*.log | flamebearer ``` <img width="1188" alt="Screen_Shot_2020-08-13_at_9_25_14_AM" src="https://user-images.githubusercontent.com/1275021/90181853-f29c8200-dd65-11ea-8163-c2e3f94a5e00.png"> Profiling showed that ~19% of the CPU time was spent inside of the `statusForModule` function. Upon further investigation, I noticed that the [normalized moduleId based on the current working directory was being generated in a loop][3], creating a new string for potentially every item in the `pending` or `ignore` config. Our app has no `ignore` config, but does have a lot of `pending` entries. Moving the `fullPathModuleId` to a variable outside the loop took the runtime from 12.94 seconds to 11.04 seconds, about ~2 seconds of savings on its own! An important thing to keep in mind for the next two steps: `statusForModule` is called _at least once per rule_ to determine if the rule should be ignored or allowed to fail (or removed from the pending file if now passing). This means that any functions ran or strings created by `statusForModule` are, in the worst case scenario, calculated `numberOfFiles * pendingRuleConfigItems * ignoreConfigItems` times if the loop doesn't find the config for the file early, as it has to search the entire array of `pending` or `ignore` items for every file. Rather than rejoining the paths, they are instead generated once and stored in a cache object. This reduces the number of strings generated and time generating the same strings over and over. Caching the result of `process.cwd` (since it seems unlikely to change once the program is started) potentially has a nice performance side effect for users on Node 10 (supported until April 2021) as [process.cwd was not cached until Node 12.2](nodejs/node#27224). Second Step Lookups for `ignore` and `pending` were separated out into different caches. This was done because while [`statusForModule` will check for a function to run][1] instead of a string, in practice this is only true for `ignore` rules. The reason `ignore` rules need the function check is that they are always a function due to being converted from [strings to functions via the micromatch module][2]. Caching the lookup of `ignore` rules reduced the need to run the functions for every rule/file. Third Step Last, but not least: `pending` rules always seem to be a list of `object`s or `string`s (as generated by `--print-pending`) and having a list that potentially has a function in it doesn't seem likely given that the recommendation from ember-template-lint is to copy/paste the output of `--print-pending`. After re-profiling, the runtime went from 12.94 seconds to 9.04 seconds on my machine. The `get-config` file no longer showed up in the profile. <img width="1116" alt="Screen Shot 2020-08-13 at 12 07 40 PM" src="https://user-images.githubusercontent.com/1275021/90181902-0647e880-dd66-11ea-9cd4-cc2778198139.png"> [1]: https://github.com/ember-template-lint/ember-template-lint/blob/5937b63bed30380b4bce0f96b19061658a176840/lib/get-config.js#L376-L377 [2]: https://github.com/ember-template-lint/ember-template-lint/blob/5937b63bed30380b4bce0f96b19061658a176840/lib/get-config.js#L280 [3]: https://github.com/ember-template-lint/ember-template-lint/blob/5937b63bed30380b4bce0f96b19061658a176840/lib/get-config.js#L374 [flamebearer]: https://github.com/mapbox/flamebearer
This commit improves the lookups of `ignore` and `pending` configs by reducing the amount of iterations required over those config properties. tl;dr: - cached the config lookups for `pending` and `ignore` to reduce the amount of times we have to re-iterate the `pending` or `ignore` arrays provided by the config. - cache string lookups for normalized moduleId paths since This was done in three steps: First Step The normalized `moduleId` was cached. Running ember-template-lint on a big codebase (668 template files) took 12.94 seconds to run. I used this command in the app's codebase directory to profile: ``` node --prof ../ember-template-lint/bin/ember-template-lint.js . ``` Next, I generated a graph with the [flamebearer][flamebearer] npm package: ``` node --prof-process --preprocess -j isolate*.log | flamebearer ``` <img width="1188" alt="Screen_Shot_2020-08-13_at_9_25_14_AM" src="https://user-images.githubusercontent.com/1275021/90181853-f29c8200-dd65-11ea-8163-c2e3f94a5e00.png"> Profiling showed that ~19% of the CPU time was spent inside of the `statusForModule` function. Upon further investigation, I noticed that the [normalized moduleId based on the current working directory was being generated in a loop][3], creating a new string for potentially every item in the `pending` or `ignore` config. Our app has no `ignore` config, but does have a lot of `pending` entries. Moving the `fullPathModuleId` to a variable outside the loop took the runtime from 12.94 seconds to 11.04 seconds, about ~2 seconds of savings on its own! An important thing to keep in mind for the next two steps: `statusForModule` is called _at least once per rule_ to determine if the rule should be ignored or allowed to fail (or removed from the pending file if now passing). This means that any functions ran or strings created by `statusForModule` are, in the worst case scenario, calculated `numberOfFiles * pendingRuleConfigItems * ignoreConfigItems` times if the loop doesn't find the config for the file early, as it has to search the entire array of `pending` or `ignore` items for every file. Rather than rejoining the paths, they are instead generated once and stored in a cache object. This reduces the number of strings generated and time generating the same strings over and over. Caching the result of `process.cwd` (since it seems unlikely to change once the program is started) potentially has a nice performance side effect for users on Node 10 (supported until April 2021) as [process.cwd was not cached until Node 12.2](nodejs/node#27224). Second Step Lookups for `ignore` and `pending` were separated out into different caches. This was done because while [`statusForModule` will check for a function to run][1] instead of a string, in practice this is only true for `ignore` rules. The reason `ignore` rules need the function check is that they are always a function due to being converted from [strings to functions via the micromatch module][2]. Caching the lookup of `ignore` rules reduced the need to run the functions for every rule/file. Third Step Last, but not least: `pending` rules always seem to be a list of `object`s or `string`s (as generated by `--print-pending`) and having a list that potentially has a function in it doesn't seem likely given that the recommendation from ember-template-lint is to copy/paste the output of `--print-pending`. After re-profiling, the runtime went from 12.94 seconds to 9.04 seconds on my machine. The `get-config` file no longer showed up in the profile. <img width="1116" alt="Screen Shot 2020-08-13 at 12 07 40 PM" src="https://user-images.githubusercontent.com/1275021/90181902-0647e880-dd66-11ea-9cd4-cc2778198139.png"> [1]: https://github.com/ember-template-lint/ember-template-lint/blob/5937b63bed30380b4bce0f96b19061658a176840/lib/get-config.js#L376-L377 [2]: https://github.com/ember-template-lint/ember-template-lint/blob/5937b63bed30380b4bce0f96b19061658a176840/lib/get-config.js#L280 [3]: https://github.com/ember-template-lint/ember-template-lint/blob/5937b63bed30380b4bce0f96b19061658a176840/lib/get-config.js#L374 [flamebearer]: https://github.com/mapbox/flamebearer
This commit improves the lookups of `ignore` and `pending` configs by reducing the amount of iterations required over those config properties. tl;dr: - cached the config lookups for `pending` and `ignore` to reduce the amount of times we have to re-iterate the `pending` or `ignore` arrays provided by the config. - cache string lookups for normalized moduleId paths since This was done in three steps: First Step The normalized `moduleId` was cached. Running ember-template-lint on a big codebase (668 template files) took 12.94 seconds to run. I used this command in the app's codebase directory to profile: ``` node --prof ../ember-template-lint/bin/ember-template-lint.js . ``` Next, I generated a graph with the [flamebearer][flamebearer] npm package: ``` node --prof-process --preprocess -j isolate*.log | flamebearer ``` <img width="1188" alt="Screen_Shot_2020-08-13_at_9_25_14_AM" src="https://user-images.githubusercontent.com/1275021/90181853-f29c8200-dd65-11ea-8163-c2e3f94a5e00.png"> Profiling showed that ~19% of the CPU time was spent inside of the `statusForModule` function. Upon further investigation, I noticed that the [normalized moduleId based on the current working directory was being generated in a loop][3], creating a new string for potentially every item in the `pending` or `ignore` config. Our app has no `ignore` config, but does have a lot of `pending` entries. Moving the `fullPathModuleId` to a variable outside the loop took the runtime from 12.94 seconds to 11.04 seconds, about ~2 seconds of savings on its own! An important thing to keep in mind for the next two steps: `statusForModule` is called _at least once per rule_ to determine if the rule should be ignored or allowed to fail (or removed from the pending file if now passing). This means that any functions ran or strings created by `statusForModule` are, in the worst case scenario, calculated `numberOfFiles * pendingRuleConfigItems * ignoreConfigItems` times if the loop doesn't find the config for the file early, as it has to search the entire array of `pending` or `ignore` items for every file. Rather than rejoining the paths, they are instead generated once and stored in a cache object. This reduces the number of strings generated and time generating the same strings over and over. Caching the result of `process.cwd` (since it seems unlikely to change once the program is started) potentially has a nice performance side effect for users on Node 10 (supported until April 2021) as [process.cwd was not cached until Node 12.2](nodejs/node#27224). Second Step Lookups for `ignore` and `pending` were separated out into different caches. This was done because while [`statusForModule` will check for a function to run][1] instead of a string, in practice this is only true for `ignore` rules. The reason `ignore` rules need the function check is that they are always a function due to being converted from [strings to functions via the micromatch module][2]. Caching the lookup of `ignore` rules reduced the need to run the functions for every rule/file. Third Step Last, but not least: `pending` rules always seem to be a list of `object`s or `string`s (as generated by `--print-pending`) and having a list that potentially has a function in it doesn't seem likely given that the recommendation from ember-template-lint is to copy/paste the output of `--print-pending`. After re-profiling, the runtime went from 12.94 seconds to 9.04 seconds on my machine. The `get-config` file no longer showed up in the profile. <img width="1116" alt="Screen Shot 2020-08-13 at 12 07 40 PM" src="https://user-images.githubusercontent.com/1275021/90181902-0647e880-dd66-11ea-9cd4-cc2778198139.png"> [1]: https://github.com/ember-template-lint/ember-template-lint/blob/5937b63bed30380b4bce0f96b19061658a176840/lib/get-config.js#L376-L377 [2]: https://github.com/ember-template-lint/ember-template-lint/blob/5937b63bed30380b4bce0f96b19061658a176840/lib/get-config.js#L280 [3]: https://github.com/ember-template-lint/ember-template-lint/blob/5937b63bed30380b4bce0f96b19061658a176840/lib/get-config.js#L374 [flamebearer]: https://github.com/mapbox/flamebearer
This commit improves the lookups of `ignore` and `pending` configs by reducing the amount of iterations required over those config properties. tl;dr: - cached the config lookups for `pending` and `ignore` to reduce the amount of times we have to re-iterate the `pending` or `ignore` arrays provided by the config. - cache string lookups for normalized moduleId paths since This was done in three steps: First Step The normalized `moduleId` was cached. Running ember-template-lint on a big codebase (668 template files) took 12.94 seconds to run. I used this command in the app's codebase directory to profile: ``` node --prof ../ember-template-lint/bin/ember-template-lint.js . ``` Next, I generated a graph with the [flamebearer][flamebearer] npm package: ``` node --prof-process --preprocess -j isolate*.log | flamebearer ``` <img width="1188" alt="Screen_Shot_2020-08-13_at_9_25_14_AM" src="https://user-images.githubusercontent.com/1275021/90181853-f29c8200-dd65-11ea-8163-c2e3f94a5e00.png"> Profiling showed that ~19% of the CPU time was spent inside of the `statusForModule` function. Upon further investigation, I noticed that the [normalized moduleId based on the current working directory was being generated in a loop][3], creating a new string for potentially every item in the `pending` or `ignore` config. Our app has no `ignore` config, but does have a lot of `pending` entries. Moving the `fullPathModuleId` to a variable outside the loop took the runtime from 12.94 seconds to 11.04 seconds, about ~2 seconds of savings on its own! An important thing to keep in mind for the next two steps: `statusForModule` is called _at least once per rule_ to determine if the rule should be ignored or allowed to fail (or removed from the pending file if now passing). This means that any functions ran or strings created by `statusForModule` are, in the worst case scenario, calculated `numberOfFiles * pendingRuleConfigItems * ignoreConfigItems` times if the loop doesn't find the config for the file early, as it has to search the entire array of `pending` or `ignore` items for every file. Rather than rejoining the paths, they are instead generated once and stored in a cache object. This reduces the number of strings generated and time generating the same strings over and over. Caching the result of `process.cwd` (since it seems unlikely to change once the program is started) potentially has a nice performance side effect for users on Node 10 (supported until April 2021) as [process.cwd was not cached until Node 12.2](nodejs/node#27224). Second Step Lookups for `ignore` and `pending` were separated out into different caches. This was done because while [`statusForModule` will check for a function to run][1] instead of a string, in practice this is only true for `ignore` rules. The reason `ignore` rules need the function check is that they are always a function due to being converted from [strings to functions via the micromatch module][2]. Caching the lookup of `ignore` rules reduced the need to run the functions for every rule/file. Third Step Last, but not least: `pending` rules always seem to be a list of `object`s or `string`s (as generated by `--print-pending`) and having a list that potentially has a function in it doesn't seem likely given that the recommendation from ember-template-lint is to copy/paste the output of `--print-pending`. After re-profiling, the runtime went from 12.94 seconds to 9.04 seconds on my machine. The `get-config` file no longer showed up in the profile. <img width="1116" alt="Screen Shot 2020-08-13 at 12 07 40 PM" src="https://user-images.githubusercontent.com/1275021/90181902-0647e880-dd66-11ea-9cd4-cc2778198139.png"> [1]: https://github.com/ember-template-lint/ember-template-lint/blob/5937b63bed30380b4bce0f96b19061658a176840/lib/get-config.js#L376-L377 [2]: https://github.com/ember-template-lint/ember-template-lint/blob/5937b63bed30380b4bce0f96b19061658a176840/lib/get-config.js#L280 [3]: https://github.com/ember-template-lint/ember-template-lint/blob/5937b63bed30380b4bce0f96b19061658a176840/lib/get-config.js#L374 [flamebearer]: https://github.com/mapbox/flamebearer
This caches the current working directory and only updates the variable
if
process.chdir()
is called.This improves the
process.chdir
performance by factor 10 andpath.resolve
byup to 66%. That again is used a lot in cjs modules.
I guess this is currently not possible since native modules could change the working directory as well, not only JS code. Or is that not a blocker / does not apply?
Do we have a way to detect if any modules are loaded or not? If so, we could have a fallback for the moment a native module is loaded.
And since
process.chdir
can not be called in workers: how does that work in combination with native modules that could change the working directory? Can we cache the working directory in case we use workers?Checklist
make -j4 test
(UNIX), orvcbuild test
(Windows) passes