[Boost-users] looking for a memory leak in an asio based application

Subject: [Boost-users] looking for a memory leak in an asio based application
From: David Medine (david.eric.medine_at_[hidden])
Date: 2017-10-05 13:35:35

I know that the answer to this question is /always/ 'no'. But is it 
possible, just possible that one can leak memory out of boost::asio?

Let me explain. I work on an open source client/server data streaming 
library for multi-modal data streaming called labstreaminglayer 
(https://github.com/sccn/labstreaminglayer). At the core of the data 
transmission architecture is boost::asio. Now, to avoid calling 
collisions, the original author of the library shipped the source code 
with a name-mangled version of boost. Over the years this has yet to be 
updated so we are still relying on boost 1.57.

The memory leak starts on the server side after a client is connected to 
it. This went undetected for years because it is very very slow and it 
depends on the sampling rate of the data stream. LSL (labstreaminglayer) 
is widely used in the neurophysiological research community (for whom 
the tool was intended), but normally in that field you are not sending 
samples faster than 500 Hz or so. But some devices are now available 
that offer relatively high sampling rates (e.g. Brain Products' 
actiCHamp can go as fast as 100kHz). The amount of memory that leaks 
seems to depend on how often data is sent, not how much data is send at 
a time. So, when the data is flowing quickly, the leak is severe.

I am 99.99% sure that the allocation of memory for all the data I am 
sending is getting deallocated correctly. I don't see any leaking memory 
according to Visual Studio's diagnostic tools, nor can Deleaker spot 
anything. However, when I look at a server process in task manager, I 
can see the memory rising. I tried the asio::io_service::work trick 
(https://stackoverflow.com/questions/32258890/c-boost-asio-async-send-to-memory-leak) 
on my io_services, but this didn't have any effect.

I have not gone through the whole pipeline yet, but I've been at this 
for two harrowing days already. So I just thought I would ask the list 
if anyone has seen anything like this before or if there has been any 
slow-leak fixes in boost (particularly asio) since 1.57 (I didn't see 
anything in the logs, but maybe I missed something). Due to the state of 
the project at this moment, it is non-trivial to update the name-mangled 
boost to another version, but it is far less trivial than the 
needle-in-a-haystack game I am playing right now. So, if someone can say 
with authority a boost update will fix everything I'll just do that and 
be happy.

It's a long-shot, I know.