Hi folks, I'm in the process of coordinating a rollout of an update from 3.2 to 3.3, and I've been coming up against some memory issues. Squid 3.3.x seems to use considerably more memory than 3.2 with the same configuration. Looking at the logs I have, 3.3 is generally using more than double the memory of 3.2 after extended use, but I think it's unbounded as memory usage eventually grows until the system kills it. The systems are 32 bit, so the memory limit imposed by the architecture is what kills it. I've written a custom malloc trace to track xmalloc() and friends, and by the looks of things, Squid is not leaking nearly as much memory as the resident shows. I can write at length about why I used this over glibc's tools, but it was basically performance driven so I can run it in a production environment. I've also run Squid with jemalloc to test if the issue is heap fragmentation, and test maximum vs actual memory usage by having unused pages returned to the operating system. In less-pronounced cases I do see the resident memory usage drop from time to time, but typically not. This makes me wonder if it is heap fragmentation or a genuine memory leak. I did find one small memory leak that was introduced by the server-first SSL bump feature in 3.3, which looks to have been fixed in 3.4, but it doesn't look to be the main cause of the memory issues I'm experiencing. This leads me to believe that the objects that are consuming all of the memory are genuinely using that memory, and are freed on shutdown. The pools in the 'mem' cache manager page don't reflect this memory usage at all, so I'm reasonably confident that it's of a type that doesn't have memory pools, or are reported inaccurately by the pools. Examples of these are the MemBuf and MemBlob objects, which contain char* members and thus variable sizes. So I'm basically at a loss as to what I can do now. Does anyone have any ideas on what I can do to squash this problem? Thanks, Nathan.
