Association Rule Learning in Hybrid Memory Systems

One of the recent computing problems to emerge is the balance between low latency dynamic random access memory (DRAM) and inexpensive storage of non-volatile memory (NVM). NVM technologies are more desirable than DRAM in the respect that they have higher memory density and lower idle power consumption making them useful for storing a bulk of the memory for a program at any given time. However, the sole use of NVM technologies cannot meet the low latency expectation that we have for fast computations. For this reason, researchers explore hybrid memory system designs that combine DRAM and NVM technology to get the benefits from each side of the spectrum. The main topic of research in this field is the page placement policy that decides when to move memory from NVM to DRAM. The difficulty in deciding a policy is to come up with a model or strategy that can effectively predict the next memory access. Due to the nature of memory accesses, it is very difficult to come up with a model that can accurately represent this. Recently, techniques in machine learning can estimate models over training datasets. Through the use of machine learning, it may be possible to accurately learn the model through training over past memory traces. The nature of our research in this paper is to test the practicality association rule learning algorithm for page prediction. The test setting for experimentation was the page ranking program executed on a non-uniform memory access (NUMA) system. We find that the implementation we propose is less efficient than program execution without a page migration policy.U of I OnlyUndergraduate senior thesis not recommended for open acces