compressor memory allocation question

Hello,

Thank you for the feedback on your use case!

You're right about the rru buffers in compressors and decompressors. I allocated it dynamically in a dev branch. It seems to work. Please test it too. See below for details.

The CRC tables are pre-computed to speed up CRC computations. Several years ago, they were not in compressor/decompressor. They were global variables. This was a kind of singleton that needed to be initialized before initializing any compressor or decompressor. It made the API simpler to put them in compressor/decompressor. It could be improved from a memory point of view, but it requires API changes. In short, it is not quick to do ;-)

About bug #1604491, there are some work in progress for performances on ARM targets. Such tiny systems are small caches, so the sizes of structures is important for performances (cache invalidation). So, there is some work in a dev branch that minimizes all structures. Please test it for your setup/use case.

The dev branch is dev_improve_perfs_decomp_on_arm on Github. It also contains the optimization that you proposed on RRU buffers. Tell me if it improves your use case.

Changelog of the branch: https://github.com/didier-barvaux/rohc/commits/dev_improve_perfs_decomp_on_arm

Regards,
Didier

This question was expired because it remained in the 'Needs information' state without activity for the last 15 days.

Hi again,

Sorry for late response.
Indeed the mentioned change with rru buffers decreases a lot the memory footprint of (de)compressor objects.

However, still the memory allocation is very high, because context's memory footprint is also too high.
One of the reason for high needed memory is for compressing lists of IPv6 extension headers, for which I am not so familiar yet.

However, by modifying some constants (eg. ROHC_LIST_ITEM_DATA_MAX) in "common\rohc_list.h", we managed to decrease size of structs list_(de)comp from 70224 to 2064.
But my questions is why to reserve memory for lists if no IPv6 support is needed? And I am referring to decompressor's context, since during compressor context initialization there is a check for IP version, and "rohc_comp_list_ipv6_new" is called only for IPv6.
Why not having the same check when init decompressor's context, when calling "rohc_decomp_list_ipv6_new" respectively?

Regards,
Evangelos

Hello,

I checked the allocated buffers for lists of IPv6 extension headers. There are indeed quite large. I did change them for the TCP profile in november, but I forgot to do the same for RFC3095 profiles. I have just did it on the dev_improve_perfs_decomp_on_arm branch. Please update your copy and tell me if it improves your situation.

There is still room for optimization for the struct list_comp. It still weights 35k. I'm adding this work on my TODO list.

Regards,
Didier

Hello,

The changes to the buffers of RRU and IPv6 extension headers are now merged in the master branch, as well as all the other performance enhancements. There will be soon released with the next major release.

Regards,
Didier Barvaux