Processing-In-Memory Primitives for Data Management (PIMPMe)

点击翻译按钮获取中文摘要

Ongoing developments in the field of memory technology open up great potential for building highly efficient data-intensive systems addressing the challenges of modern applications. At the same time, it requires to rethink traditional assumptions to fully exploit this technology.In this project we plan to investigate the benefit of memory-centric computing paradigms: we will conduct research on offloading computation to memory in order to reduce the amount of data to be transferred between memory and CPU and in this way increase bandwidth, to reduce the CPU load for computation, and in the end, help to reduce the energy consumption of modern IT systems.We plan to particularly conduct research in three areas where the overarching goal is to develop a framework of primitives as building blocks for typical data management tasks such as query processing, logging for recovery, and transaction management. The three areas are Processing-in-Memory (PIM) where operations are executed directly in memory, Processing-near-Memory (PNM) bringing processing closer to memory and relieving the CPU, e.g. by FPGA-based accelerator cards with device memory, and Transactional Memory to provide the abstraction of atomicity for critical code sections.The project has three main objectives: first, we plan to design and implement PIM and PNM primitives for query processing on graph data, index management, and log management. Second, we want to improve concurrency control in databases by exploiting memory operations and transactional memory. Third, we investigate strategies to map database transactions/queries to these primitives by building a comprehensive framework of primitives and evaluating them in the context of graph database processing.