Sorry for bumping this thread up again, but I'm still fighting with this problem.
I have done some more testing, used a different configuration file, changed a little bit the schema and I have some result that maybe someone here can help me reading.
1) changed the configuration file to MySQL 5.6 default: no relevant improvement
2) disabled partitioning (I have 5 tables with 2000 partitions each), and seems that disabling partitioning helps in some way. But still I have not clear the reason, and I cannot eliminate partitioning in production
3) reduced the number of partitions on tables: apparently the behavior is the same, the growth curve is less ripid but still I have some.
4) inserted some random sleep between queries: this produced a good flat memory picture, but, of course this cannot be done in production.
5) run some basic test running mysql under valgrind for leak detection:
http://pastebin.com/W3ZMAGY4
And massif:
http://pastebin.com/LweQHn3M .
These tests are executed with MySQL default configuration for a small period of time with a component that makes several queries in parallel.
If I'm reading the output correctly, no relevant leak is found but I have a huge memory consumption per connection (500 MB each connection?), which possibly grows over time.
Is that correct? Any idea why it grows so much?