Title: Learning Slab Classes to Alleviate Memory Holes in Memcached

Abstract: We consider the problem of memory holes in slab allocators, where an item entered into memory occupies more memory than it actually requires due to a difference between the nearest larger slab class size and the size of the entered item. We solve this problem by using a greedy algorithm that analyses the pattern of the sizes of items previously entered into the memory and accordingly re-configuring the default slab classes to better suit the learned traffic pattern to minimize memory holes. Using this approach for a consistent data pattern, in our findings, has yielded significant reductions in memory wastage. We consider Memcached as it is one of the most widely used implementations of slab allocators today, and has native support to reconfigure its default slab classes.