significantly higher memory consumption and binary size with mbed 5.1

[mfiore02]

We are currently developing a platform using the STM32L151CC processor - 32kB of RAM and 256kB of flash. We added platform support to mbed 2.0 during the transition to 5.x and are now adding support to 5.x so we can make a pull request and have our platform enabled.

We've noticed significantly higher memory consumption in mbed 5.x vs mbed 2.0. This is a problem for us because our application (an AT parser and a LoRa stack) currently can't run due to the system running out of memory during initialization.

We've done some testing with a similar platform to ours, NUCLEO-L152RE. We took the same main.cpp and built it with mbed-os revision 2244, mbed 125 & rtos 121, and mbed 121 & rtos 117. All of these builds were using the mbed online compiler. The main from the app is below:

  #include "mbed.h"
  #include "rtos.h" // only in the mbed 2.0 version

  void func(void const*) {
      while (true) {
          osDelay(100);
      }
  }

  int main() {
      uint8_t* mem_buf = NULL;
      uint32_t mem_size = 0; 
      Thread t(func);

      while (true) {
          // malloc until we die
          mem_buf = new uint8_t[256];
          if (! mem_buf) {
              printf("malloc failed: %lu\r\n", mem_size);
          } else {
              mem_size += 256;
              printf("%lu\r\n", mem_size);
          }
      }
  }

Our results are as follows:

with mbed revision 121 mbed-rtos revision 117
binary size: 20kB
free memory: 74752

with mbed revision 125 mbed-rtos revision 121
binary size: 29K
free memory: 66816

with mbed-os revision 2244
binary size: 37kB
free memory: 65792

When I compile our LoRa stack (offline) for our new platform with mbed 2.0, I'm able to allocate over 13kB before the system runs out of memory. When I compile (offline) with mbed 5.1, I'm only able to allocate 2kB.

I'm hoping there are some switches in mbed 5.x that I can flip to reduce memory consumption. Reducing the consumption of our application and stack will be time-consuming and is considered a last resort.

Any ideas or insight appreciated.

[sg-]

Some information about the different ROM sizes:

• NUCLEO-L152RE with mbed revision 121 mbed-rtos revision 117 used uARM as the default toolchain. Baseline
• NUCLEO-L152RE with mbed revision 125 mbed-rtos revision 121 uses ARM as the default toolchain. ~9k increase
• with mbed-os revision 2244 has some test harness code lingering that is getting linked in. This will be fixed. ~8k increase that will be reverted along with a few more optimizations

I'm able to allocate over 13kB before the system runs out of memory. When I compile (offline) with mbed 5.1, I'm only able to allocate 2kB.

The default stack size has increased from 128x4 to 512x4 so this will take a chunk of that as well as a fixed size main stack. Also, GCC has some allocation for buffering in the stdlib. Lots of work ongoing right now around memory optimizations. Thanks for the detailed report. Any more finding, please report them here!

[ciarmcom]

ARM Internal Ref: IOTMORF-452

[janjongboom]

@mfiore02 There are three patches in the work which will bring memory consumption back to mbed 2.0-like levels (4.8K decrease in RAM for blinky), plus we're making the stack size configurable in #2646 so you can reclaim some memory there as well.

We'll be publishing a blog post with some background and hard numbers when merged, but I'll send you a draft.

[mfiore02]

@janjongboom that's great news! We've manually tweaked stack sizes for our platform for now, but it's good to know there's a global resolution coming!

[nuket]

I've got similar out of memory issues using mbed 5.1 with an nRF51-DK (256KB flash / 32KB RAM).

Checking one app with a similar "malloc(64) until it dies" strategy, I get the following mbed_stats_heap_t output:

• Start: Heap stats - alloc_cnt: 6, alloc_fail_cnt: 0, current_size: 5284, max_size: 5284, total_size: 5284
• Stop: Heap stats - alloc_cnt: 26, alloc_fail_cnt: 0, current_size: 6564, max_size: 6564, total_size: 6564

Meaning that I only have about 1K of RAM left to use in my program.

On another app I'm working on, the initial allocations succeed at boot time and tell me:

• Heap stats - alloc_cnt: 11, alloc_fail_cnt: 0, current_size: 6872, max_size: 6872, total_size: 6872

But the minute I try one more new of any size, the system jumps to HardFault and prints Operator new out of memory on the serial debug port.

Hope these memory reduction patches make it to the mainline soon!

[janjongboom]

@nuket Yeah, I've seen the same issues on the nRF51, but with these patches we managed to fit BLE + mbed TLS + RTOS + application into 32K again, similar profile as on mbed 2.

[nuket]

@janjongboom This sounds pretty excellent.

[nuket]

@janjongboom Are the patches associated with this stable / getting closer to being merged?

[janjongboom]

ping @pan-

[pan-]

I've made three different PR to solve some issues we currently have with the memory consumption: #2715 #2741 and #2745.
Hopefully they will be merged soon but it will not fix everything,

Here is some good practices to optimize the application for size:

• Avoid the standard IO subsystem (apply Allow the trace output by mbed error to be conditional of NDEBUG. #2715, Add configuration for IO flushing during exit. #2741 and compile with NDEBUG), it saves 2K of RAM and 14K of flash with ARMCC (online compiler).
• If you have a lot of static global objects, apply Disable global objects destruction #2745 to reduce the number of memory allocations needed for destruction.
• Allocate statically as much as you can.
• Allocate small objects using a pool allocator to avoid fragmentation of the main heap.
• adjust the parameters of the os to your application: All the parameters of the OS are defined in this file. You can override then by putting a file named mbed_app.json at the root of your application. This file contains the application configuration (more details here). Every define can be overridden in the macros list:

{ 
  "macros": [ 
    "NDEBUG=1",
    "OS_TASKCNT=XXX",
    "..."
  ]
}

the main macros to override are:

• OS_TASKCNT: The number of task running in parallel in the application.
• OS_IDLESTKSIZE: The size of the stack for the idle task, in words of 4 bytes. It can be shrinked by default to 32.
• OS_STKSIZE Can be shrunk to 1 if you drive provide manually the buffer which will be used by each threads stack.
• OS_MAINSTKSIZE: The stack of the main thread in words of 4 bytes.
• OS_TIMERS: 1 or 0, enable or disable the timer thread.

[mfiore02]

@pan- I see these fixes have been merged into master branch. Looks like we should expect to see them in mbed-os-5.2?

Is there a release date for mbed-os-5.2?