It turns out LRU has two shortcomings when used as a page replacement algorithm. First, an exact LRU implementation is too costly in this context. Second, the memory manager needs to account for frequency as well, so that a large file read doesn’t evict the entire cache. Therefore, Linux uses a more sophisticated algorithm than LRU; and that algorithm doesn’t play along well with the workload we just described.

via CitusDB: Scalable Analytics Database – Linux memory manager and your big data.