Making Cold Data Identification Efficient in Non-volatile Memory Systems

Abstract

Non-volatile memory is emerging as a promising candidate for building efficient data-intensive OLTP systems, due to its advantages in high area density and low energy consumption. Systems now are able to store large datasets in main memory. Because OLTP workloads typically exhibit skew access patterns, the system must maintain an eviction order policy to move the cold data to the economical secondary storage. Existing cold data identification schemes generally employ the linear lists to track the least recently used data. However frequently update cost in these schemes is extremely high which is unsuitable to identify cold data from large scale of memory-resident data. We propose an efficient cold data identification scheme named eLRU. eLRU is a trie-based LRU which is able to fast track billions of tuples. We implemented our eLRU proposal and performed a series of experiments across a range of database sizes, workload skews and read/write mixes. Our results show that eLRU has a 2$\times$–4$\times$ performance advantage over the current LRU-based cold data identification schemes.