On Sat, Jun 25, 2016 at 05:46:47PM +0200, Sven Eckelmann wrote:
On Saturday 25 June 2016 16:44:06 Sven Eckelmann wrote:
The translation table (global, local) is usually the part of batman-adv which has the most dynamical allocated objects. Most of them (tt_local_entry, tt_global_entry, tt_orig_list_entry, tt_change_node, tt_req_node, tt_roam_node) are equally sized. So it makes sense to have them allocated from a kmem_cache for each type.
This approach allowed a small wireless router (TP-Link TL-841NDv8; SLUB allocator) to store 34% more translation table entries compared to the current implementation.
[1] https://open-mesh.org/projects/batman-adv/wiki/Kmalloc-kmem-cache-tests
Reported-by: Linus Lüssing linus.luessing@c0d3.blue Signed-off-by: Sven Eckelmann sven@narfation.org
Here is the changelog (since the RFC):
v2:
- use kmem_cache_zalloc instead of kmem_cache_alloc + memset as suggested by Linus
- add cache batadv_tt_change_cache, batadv_tt_req_cache, batadv_tt_roam_cache
- Added Reported-by
- rewrote commit message
- summarize+link the results from Linus' tests
Looks good. Do you have any typical numbers regarding tt_change_cache and tt_req_cache sizes for specific setups?
Just wondering whether it is worth it to have extra caches for these two. (if you don't have any numbers, I might have a look at /proc/slabinfo with your patch for some Freifunk network(s) later today)