On Oct 16, 3:15 pm, Marc Bollinger <[email protected]> wrote:
> We've been testing a write-heavy scenario (hence the get/set > discrepancy) using the Jellycan build of 1.2.6 on Windows Server 2003. > There's a staging server throwing live data at the memcached server > (items are roughly 2-300 bytes apiece), and are seeing the number of > curr_items stagnate at exactly 22539, at which point everything else > is evicted. I haven't perused your stats (and there are a couple more required to really understand), but this would be the expected behavior based on what you're saying. You've used chunks up to 11k and memory has been allocated for these things. Once the memory has been allocated and spread out across slabs, evictions are the only way to get more items stored. We have more and better stats for visibility and options to make it easier to tune it more for your application in the 1.4 series (which we're very actively working towards bringing to Windows users right now).
