snappier place startup time

[Nate Griswold]

Is there any way to make places startup faster? Even if i do an explicit round trip using place-channel-put and place-channel-get on both sides, it takes on the order of centiseconds for near empty places to start up.

My program requires the threads for a couple places to be set up before it can operate, so this impacts my startup time by quite a bit.

I have one place that has a very simple module and one place with a more complicated module. Is there some sequence that i should do things in for the minimal startup time? It seems nothing i do really helps much.

Nate

[Nate Griswold]

Oh, and i'm still on Racket 7.8. Don't know if 7.9 addresses anything.

Nate

[Matthew Flatt]

The bottleneck for place startup is loading modules into the new place,
including modules like `racket/base`.

For example,

(place-wait (dynamic-place 'racket 'void))

takes around 200ms on my machine, while

(place-wait (dynamic-place 'racket/base 'void))

takes around 30ms and

(place-wait (dynamic-place 'racket/kernel 'void))

takes around 10ms.

It sounds like you're already aware that the complexity of the module
loaded into a place matters, though. Beyond using a minimal set of
modules, I don't have any way to make place startup faster.

Matthew

> --
> You received this message because you are subscribed to the Google Groups
> "Racket Users" group.
> To unsubscribe from this group and stop receiving emails from it, send an
> email to racket-users...@googlegroups.com.
> To view this discussion on the web visit
> https://groups.google.com/d/msgid/racket-users/CAM-xLPpvfCHHDDpfNmuTWQOyfYfEJ7v
> m1c_dS7nj3FxaEFVm2Q%40mail.gmail.com.

[Nate Griswold]

Oops, i am having some issues with not getting to the list from my other email address. Here is a reply i sent for the record.

---

Thank you, Matthew.

The following code takes around 250ms on my machine. Any idea why? I was expecting it to be fast since the module is based on racket/base.

#lang racket/base

(require syntax/location)

(require racket/place)

    

(module test racket/base

 (provide place-main)

racket

 (define (place-main pch)

  (void)))

(time (place-wait (dynamic-place (quote-module-path test) 'place-main)))

Nate

On Tue, Nov 24, 2020 at 8:35 AM Nathaniel W Griswold <na...@manicmind.earth> wrote:

Thank you, Matthew.

The following code takes around 250ms on my machine. Any idea why? I was expecting it to be fast since the module is based on racket/base.

#lang racket/base

(require syntax/location)

(require racket/place)

    

(module test racket/base

 (provide place-main)

racket

 (define (place-main pch)

  (void)))

(time (place-wait (dynamic-place (quote-module-path test) 'place-main)))

Nate

[Nathaniel W Griswold]

Thank you, Matthew.

The following code takes around 250ms on my machine. Any idea why? I was expecting it to be fast since the module is based on racket/base.

#lang racket/base

(require syntax/location)

(require racket/place)

    

(module test racket/base

 (provide place-main)

racket

 (define (place-main pch)

  (void)))

(time (place-wait (dynamic-place (quote-module-path test) 'place-main)))

Nate

On Nov 24, 2020, at 8:16 AM, Matthew Flatt <mfl...@cs.utah.edu> wrote:

[Sam Tobin-Hochstadt]

Almost certainly the problem is expansion time. If I run that program
on my machine, it takes about 200 ms. But if I compile the file to zo
first with `raco make`, then it takes about 40 ms, basically identical
to `racket/base`.

Sam

> To view this discussion on the web visit https://groups.google.com/d/msgid/racket-users/CAM-xLPqtJrem4j%3DUi3fbrduoahsXCNNA2JPuB0Tt9dissiu5KA%40mail.gmail.com.

[Nate Griswold]

Awesome, thanks!

Nate

[Matthew Flatt]

Just to elaborate a little more:

The difference is because the `test` submodule can be loaded
independently from the compiled form. Loading the submodule from source
requires loading the enclosing module, too (which depends on
`racket/place` and more).

> https://groups.google.com/d/msgid/racket-users/CAM-xLPqVgEBvrRzjU7%3DX_h3Wy_YUH
> 11G6CX5%2BKjSct26pi3oEA%40mail.gmail.com.

[Nate Griswold]

Oh, interesting. So compilation breaks the submodule out from the modules if possible?

So anyway, it sounds like breaking my modules out into separate files will improve performance in most cases.

Unfortunately, i need racket/place in the module that is my startup bottleneck. If i modify the previous program to require racket/place and compile, it takes around 180ms.

This is about what i can expect for a module that requires racket/place, then?

Nate

[Robby Findler]

DrRacket thinks that there are no references to a number of the requires in racket/place, including racket/fixnum, racket/flonum, racket/vector, and racket/runtime-path. Not sure if that's an error on DrRacket's part (and I don't see why those would be needed for their effects).

Also, the only use of racket/match seems to be this, which seems simple to rewrite out.

    (match name
      [(? symbol?) `(submod (quote ,name) ,submod-name)]
      [(? path?) `(submod ,name ,submod-name)]
      [`(,p ,s ...) `(submod ,(if (symbol? p) `(quote ,p) p) ,@s ,submod-name)])

Robby

To view this discussion on the web visit https://groups.google.com/d/msgid/racket-users/CAM-xLPoLD6N%3Dwep%2Bed5UGduA7mZ-SaYxFJ9VbqM78CppMG3m2w%40mail.gmail.com.

[Nathaniel W Griswold]

Cool. If this is indeed the case it might be nice for someone (maybe me) to cut it down, since any nontrivial place will of course require racket/place and that is kind of a long time.

Nate

[Robby Findler]

I didn't check to see if removing those has a significant performance effect, but the remaining requires seem pretty minimal.

Robby

To view this discussion on the web visit https://groups.google.com/d/msgid/racket-users/315E2C2B-F314-4044-8ACC-1D66288BCCFB%40manicmind.earth.

[Nathaniel W Griswold]

I checked into it a bit.

racket/fixnum, racket/flonum, and racket/vector are needed by “private/th-place.rkt”, which is required by racket/place. Not sure why DrRacket is saying that it’s not needed.

racket/runtime-path does not appear to be needed.

I tried removing racket/runtime-path and racket/match but didn’t see any performance gains. It appears the delay is elsewhere.

Nate

On Nov 24, 2020, at 9:52 AM, Robby Findler <ro...@cs.northwestern.edu> wrote:

[Nathaniel W Griswold]

Actually, it cuts about 20-25ms off of a single import. Down from 185ms to 165ms for me. 50ms off my startup time of my app on average, since i basically stack the import twice and sync on the place being ready.

Might be worth including and seeing if there’s anything else that can be shaved off.

Nate

[Nathaniel W Griswold]

I seem to remember there being some global namespace. Since every reasonable place will require racket/place, might it be possible to make the racket/place import a special case and stick it in the global space, to improve place setup time? It would be nice to be able to only set up racket/place one time instead of once for each place.

Nate

[Nathaniel W Griswold]

I noticed i am using the pl- functions so I replaced th-place.rkt with a stub and saw more time shaved off, this time about 15ms for each racket/place import.

Under what circumstances is th-place used instead of '#%place and needed?

Nate

[Sam Tobin-Hochstadt]

th-place is used if places are not enabled when Racket is built (this is the default on some platforms). 

I'm making progress on shrinking this, hopefully I'll have a patch done soon. 

One thing to note is that '#%place can be required directly and will have almost no start-up cost. 

Sam

Sam

To view this discussion on the web visit https://groups.google.com/d/msgid/racket-users/7AE1B364-8DE1-435C-8AA2-EBCFE078040D%40manicmind.earth.

[Robby Findler]

On Tue, Nov 24, 2020 at 12:09 PM Nathaniel W Griswold <na...@manicmind.earth> wrote:

racket/fixnum, racket/flonum, and racket/vector are needed by “private/th-place.rkt”, which is required by racket/place. Not sure why DrRacket is saying that it’s not needed.

Ah, sorry: DrRacket was merely saying that the exports of those modules weren't used by the racket/place module.

Robby

[Sam Tobin-Hochstadt]

Here's an in-progress PR: https://github.com/racket/racket/pull/3518

With this, your simple test takes about 150ms with `racket/place` and
50ms with `racket/place/dynamic`. For comparison, just having the
submodule depend on `racket/base` gives a time of about 42 ms, and
just `racket/kernel` is about 12ms.

Sam

[na...@manicmind.earth]

Awesome. Also I didn’t know you could include ‘#%place like that.

Nate

> On Nov 25, 2020, at 9:41 PM, Sam Tobin-Hochstadt <sa...@cs.indiana.edu> wrote:
>
> Here's an in-progress PR: https://github.com/racket/racket/pull/3518