WASM built with USING_PTHREADS=1 can't go beyond 1GB

[Prashanth Nethi]

Hi,

I am currently building WASM with the following flags, to enable PThreads in Wasm.

-s USING_PTHREADS=1 -s INITIAL_MEMORY=1999MB -s MAXIMUM_MEMORY=2GB. 

This works wonderfully for our use cases! In fact we are able to get 2x performance in some cases!

When I checked the max memory that the Wasm could use, with PThreads enabled, it got capped at 1 GB. I am seeing that when the WASM is built with ALLOW_MEMORY_GROWTH, the Wasm can use upto 2GB. I know that ALLOW_MEMORY_GROWTH with USE_PTHREADS is discouraged so can't look at that as a possible solution.

Is there anyway I can get Wasm to use 2GB (or even potentially 4GB in the future) with PThreads enabled? Is it that I am missing using some configuration options? 

I am really hoping there is a way to increase the WASM cap to 2GB, as using PThreads, solves our use cases in a big way.

Thanks,

Prashanth Nethi

[Alon Zakai]

I think you can do any number up to 2GB, including 2GB - 64Kb. So the limit isn't 1GB, unless you see that on some specific browser? Could be a bug.

It should soon be possible to do up to 4GB for the initial memory (without growth), thanks to a spec change, https://github.com/WebAssembly/spec/pull/1174

--
You received this message because you are subscribed to the Google Groups "emscripten-discuss" group.
To unsubscribe from this group and stop receiving emails from it, send an email to emscripten-disc...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/emscripten-discuss/730a6796-5b14-4a9e-a1d8-298415c67cd1n%40googlegroups.com.

[Prashanth Nethi]

Thanks Alon! So here is something very weird. I could get the memory usage go all the way to 2GB when I changed my testing code. This was my original test code. So basically I was just adding elements to std::vector infinitely.

class TestClass{

 private:

  int t = 0;

};

struct Data {

 int t;

 TestClass obj;

};

typedef std::vector<Data> Vec;

Vec someVec;

using namespace std;

int main() {

 printf("hello, world!\n");

 while(1){

  Data data;

  someVec.push_back(data);

 }

 return 0;

}

With this code, the WASM memory was going all the way to 1GB.

But when I changed the code to this, where I am writing some value after acquiring memory, then I am able to see the memory usage go all the way up to 2 GB. Could this be a bug? I am on emscripten 2.0. 

int main() {

  printf("hello, world!\n");

  char *p = nullptr;

  int byteSize = 50 * 1024 * 1024;

  while(1){

        p = new char(byteSize);

        p[byteSize] = 20;

  }   

  return 0;

}

Also It is very encouraging to see that 4GB is considered for PThreads as well! Thanks.

One follow up question. May be a dumb one. What could be the potential problems with ALLOW_MEMORY_GROWTH missing in PThreads mode? I see that when the Wasm is instantiated, the overall memory that the Chrome tab was taking was similar to the one taken by the WASM built with ALLOW_MEMORY_GROWTH. Is it that, we will not be able to instantiate WASM on low end devices if built with ALLOW_MEMORY_GROWTH=0?

Greatly appreciate your help Alon!

Thanks,

Prashanth Nethi

[Alon Zakai]

My guess is that's because of the behavior of std::vector and how it resizes. Over those appends it will malloc and free repeatedly and that may cause fragmentation that prevents a final larger size, which must be a single contiguous region. The second version allocates many smaller ones, not a single contiguous region.

- Alon

To view this discussion on the web visit https://groups.google.com/d/msgid/emscripten-discuss/86a9fc74-2036-4749-8212-29f6802615d0n%40googlegroups.com.

[Prashanth Nethi]

Thanks Alon! That explains it! Yeah I should have thought a little deeper.

I am just posting my follow up question in case you did not get a chance to look at it.

One follow up question. May be a dumb one. What could be the potential problems with ALLOW_MEMORY_GROWTH missing in PThreads mode? I see that when the Wasm is instantiated, the overall memory that the Chrome tab was taking was similar to the one taken by the WASM built with ALLOW_MEMORY_GROWTH. Is it that, we will not be able to instantiate WASM on low end devices if built with ALLOW_MEMORY_GROWTH=0?

Regards,

Prashanth Nethi

[Alon Zakai]

On Thu, Aug 27, 2020 at 8:47 AM Prashanth Nethi <prashan...@gmail.com> wrote:

Thanks Alon! That explains it! Yeah I should have thought a little deeper.

I am just posting my follow up question in case you did not get a chance to look at it.

One follow up question. May be a dumb one. What could be the potential problems with ALLOW_MEMORY_GROWTH missing in PThreads mode? I see that when the Wasm is instantiated, the overall memory that the Chrome tab was taking was similar to the one taken by the WASM built with ALLOW_MEMORY_GROWTH. Is it that, we will not be able to instantiate WASM on low end devices if built with ALLOW_MEMORY_GROWTH=0?

I'm not sure what you're asking here?

In general, not having memory growth enabled means that memory can't grow. So if you need more than the initial value, the program will hit a problem. I don't think there's anything special to pthreads in that case. (The reverse, having growth *enabled*, does have downsides for pthreads as the JS use of memory becomes somewhat slower.)

To view this discussion on the web visit https://groups.google.com/d/msgid/emscripten-discuss/9ddd4487-ff48-4b05-a138-900103ec2a4dn%40googlegroups.com.

[Prashanth Nethi]

My bad Alon! I will try to elaborate the scenario.I am trying to understand the implications of switching off ALLOW_MEMORY_GROWTH in our project. (which would be the case if we want  PTHREADS enabled).

The question is around, what if we set INITIAL_VALUE value to max value (2GB).  Does that mean when WASM is instantiated with INITIAL_VALUE=2048MB, 2GB is reserved right upfront, even if not required right away? If yes, does that mean this will reduce the usable JS heap size (by 2GB), right from the beginning?

When I instantiate WASM (in my test app) with an INITIAL_VALUE=2000MB and check for the memory that specific webpage is taking, I see that page does not take 2GB but a lot lesser. It is when I start acquiring more memory, the memory usage goes up until it hits the 2GB limit. Surprisingly this is the same behaviour I see with  ALLOW_MEMORY_GROWTH =1, USE_PTHREADS=0 (i.e. with PThreads disabled). So trying to understand the dynamics and come up with the recommendation on whether to enable or not enable PTHREADS in our app. FYI. The app has the requirement to load on various browsers and devices, with Chrome and Chromebook being our majority targets.

Regards,

Prashanth Nethi

[Alon Zakai]

On Thu, Aug 27, 2020 at 10:28 AM Prashanth Nethi <prashan...@gmail.com> wrote:

My bad Alon! I will try to elaborate the scenario.I am trying to understand the implications of switching off ALLOW_MEMORY_GROWTH in our project. (which would be the case if we want  PTHREADS enabled).

The question is around, what if we set INITIAL_VALUE value to max value (2GB).  Does that mean when WASM is instantiated with INITIAL_VALUE=2048MB, 2GB is reserved right upfront, even if not required right away? If yes, does that mean this will reduce the usable JS heap size (by 2GB), right from the beginning?

Yes, exactly. An initial value of X means X is allocated right from the start. Yes, this reduces available memory for other things, which can have downsides.

When I instantiate WASM (in my test app) with an INITIAL_VALUE=2000MB and check for the memory that specific webpage is taking, I see that page does not take 2GB but a lot lesser.

How are you measuring that?

It's possible the browser allocates that memory via calloc() or such, and maybe the OS doesn't actually use any physical memory until those pages are touched, though. So maybe only virtual memory is used initially. (But even that can cause problems on 32 bit due to address space limits.)

Measuring via browser devtools should report the full 2GB is used immediately.

 

To view this discussion on the web visit https://groups.google.com/d/msgid/emscripten-discuss/f197c8df-ba90-4f68-9018-3590303302ean%40googlegroups.com.

[Prashanth Nethi]

Thanks for the information Alon! That is exactly the information I wanted. Your theory of deferred memory usage pattern might be the reason for browsers reporting used memory differently.

It is unfortunate that we will not be able to use PThreads in our main Wasm because of this limitation, as we have lot of JS running alongside Wasm. Any rough timeline on when we can expect ALLOW_MEMORY_GROWTH to work with PTHREADS?

Also about checking the memory usage in devtools, I am using Chrome's task manager as well as Activity Monitor (both on Mac) to check the webpage's memory footprint. At both the places, the 2GB reserved memory is not getting reflected. Maybe I am missing on checking other relevant fields. But that should be fine, as I got the required information from you.

Appreciate the help Alon!

Thanks,

Prashanth Nethi

[Alon Zakai]

On Fri, Aug 28, 2020 at 6:17 AM Prashanth Nethi <prashan...@gmail.com> wrote:

Thanks for the information Alon! That is exactly the information I wanted. Your theory of deferred memory usage pattern might be the reason for browsers reporting used memory differently.

It is unfortunate that we will not be able to use PThreads in our main Wasm because of this limitation, as we have lot of JS running alongside Wasm. Any rough timeline on when we can expect ALLOW_MEMORY_GROWTH to work with PTHREADS?

It already works, but memory access from JS is somewhat slower. In most cases you won't notice that, though - unless you've already tested and see overhead? If so that could be useful to mention to the standards bodies that are considering a spec change that could improve this, it could increase the priority.

- Alon

To view this discussion on the web visit https://groups.google.com/d/msgid/emscripten-discuss/a03fe2fa-6203-4342-8e1f-49edaeeac272n%40googlegroups.com.

[Prashanth Nethi]

That's interesting to know that ALLOW_MEMORY_GROWTH already works with PThreads. I will try to do some tests and see how that goes.

Could you tell if there is any performance impact on writing strings (using Module._malloc()) or binary data to the heap, with ALLOW_MEMORY_GROWTH on PThreads. Also I am assuming there is no/very less impact of going from JS to C++ via ccall or embind. Please correct me if I am wrong.

Also upon enabling ALLOW_MEMORY_GROWTH on our project, there are lots of warning being thrown up in the console. Is there any switch that I could use to disable this warning?

root:WARNING: USE_PTHREADS + ALLOW_MEMORY_GROWTH may run non-wasm code slowly, see https://github.com/WebAssembly/design/issues/1271

Thanks,

Prashanth

[Alon Zakai]

On Tue, Sep 1, 2020 at 9:20 AM Prashanth Nethi <prashan...@gmail.com> wrote:

That's interesting to know that ALLOW_MEMORY_GROWTH already works with PThreads. I will try to do some tests and see how that goes.

Could you tell if there is any performance impact on writing strings (using Module._malloc()) or binary data to the heap, with ALLOW_MEMORY_GROWTH on PThreads. Also I am assuming there is no/very less impact of going from JS to C++ via ccall or embind. Please correct me if I am wrong.

Writing data from JS to memory can be slower, including writing strings. But hopefully not by much. Nothing else should be affected.

 

Also upon enabling ALLOW_MEMORY_GROWTH on our project, there are lots of warning being thrown up in the console. Is there any switch that I could use to disable this warning?

root:WARNING: USE_PTHREADS + ALLOW_MEMORY_GROWTH may run non-wasm code slowly, see https://github.com/WebAssembly/design/issues/1271

We could add a flag to allow disabling that warning. I think it could use diagnostics.warning() in emcc.py instead of just printing it unconditionally. A PR would be welcome, or an issue.

 

To view this discussion on the web visit https://groups.google.com/d/msgid/emscripten-discuss/4175ad68-aab5-4190-9463-749a4fec26den%40googlegroups.com.

[Prashanth Nethi]

Thanks for the information Alon. I would test out and share the results.

Also I will try to see if I can put a PR for conditionally disabling the ALLOW_MEMORY_GROWTH + USE_PTHREADS warning.

Regards,

Prashanth

[Prashanth Nethi]

Hi Alon,

I have raised a PR for disabling the USE_PTHREADS + ALLOW_MEMORY_GROWTH warning conditionally.

https://github.com/emscripten-core/emscripten/pull/12103

Could you please have a look? 

Regards,

Prashanth Nethi