Hi,
I'm using HDF5 1.8.15-patch1 whose library version is 10.0.1 (C++). OS is a
Debian Wheezy.
I have a very simple test code that is trying to find a possible memory leak
I'm observing in the code.
(...)
H5File file(...);
for (int i = 0; i < 1000; i++)
{
DataSet ds = file.openDataSet("MyDS"); // (1) intentionally kept inside the
loop [see note below]
DataSpace dataspace(ds.getSpace()); // (2)
// doSomething
dataspace.close(); // (3)
ds.close(); // (4)
sleep_for(milliseconds(20));
}
This program runs for 20sec and I keep seeing physical memory increasing every
second. If line (2) is commented out, memory footprint stays at the same level
throughout the whole program execution.
Notes:
- Version 1.8.16 (C++ API) is removing memory leaks as per this
release doc
(https://www.hdfgroup.org/ftp/HDF5/current/src/unpacked/release_docs/RELEASE.txt).
Would this update also fix the getSpace() call?
- There is an old post in StackOverflow that shows exactly this
problem I'm having
(http://stackoverflow.com/questions/13443689/why-does-this-code-leak-simple-codesnippet).
But there is no good answer for that.
- I was originally writing data to a dataset in that loop above. Then,
I read this page (https://www.hdfgroup.org/HDF5/faq/perfissues.html, its third
bullet in section Open Objects) and noticed that I was not closing the
dataspace. But that didn't bring any positive effect unfortunately. Memory
still keeps going up.
- If I call file.getObjCount() inside that loop, I see it returns 1.
The call file.getObjCount(H5F_OBJ_FILE) also returns 1, meaning that it's only
the file that is opened.
- The code snippet above is just an example and what is inside the
loop (not the loop itself) is actually inside a wrapper that abstracts the HDF5
library from the client. So, I don't know if the wrapper user will use that
inside a loop or not. That's why I want to intentionally keep lines (1, 2, 3
and 4) inside the loop.
Could anyone please shed some light on this?
Thanks!
Carlos Rocha
_______________________________________________
Hdf-forum is for HDF software users discussion.
[email protected]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5
