Energy-Efficient, Robust and Secure Data Storage Using Data Mining and Analysis
Principal Supervisor: Dr. Georgios Karakonstantis
+ Project Description
The explosive growth of generated data by the increasing number of connected devices and the resultant need for more memory capacity is driving the aggressive scaling of Dynamic Random- Access Memories (DRAM). However, aggressive DRAM scaling is hampered by the need of periodic refresh cycles to retain the stored data, the frequency of which is conventionally being determined by the worst case retention time of the weakest cell. Such an approach might guarantee error free storage but its viability is questionable due to the incurred large power and performance overhead that may reach up-to 25-50% in future densities.
The aim of this project is to reveal new refresh-rates and voltages that could be used in DRAMs within state-of-the-art servers. To this end, a ‘smart’ automatic code generation framework based on machine learning will be developed for the characterization of DDR3 and DDR4 memories under reduced refresh rates and voltages as well as dynamically varying environmental conditions (e.g. temperature). The developed framework will help accelerate the characterization process and the identification of weak cells, while considering several data, access and temperature dependent factors. A follow-up data analysis stage will target the correlation of the observed behaviour with typical key performance counters, which could be used during the optimization of the DRAM and server operation. Apart from the cells prone to failures the framework will target the identification of cells and memory areas sensitive to security attacks.
Finally, a data-mining based scheme for identifying any unexpected DRAM behaviour and for correcting or at least minimizing the impact of any fault on system functionality will be developed and evaluated using a set of server related workloads. Apart from DDR3 and DDR4 DRAMs the project will also target the characterization of the emerging hybrid DRAM and Non-Volatile Memory technologies (i.e. NVDIMM) under various conditions and will investigate the robustness of a system that combines both technologies for storing data selectively depending on their ‘criticality’.
The project will built upon the server focused experimental infrastructure that has already been developed in
the Data Science and Scalable Computing Centre of ECIT.
The PhD studentship will be based at the DSSC Centre of the Queen’s Global Research Institute of Electronics, Communications and Information Technology (ECIT).
edia data in a real-time environment. This will involve undertaking research in the area of information retrieval, machine learning and natural language processing.
+ How to Apply
Applicants should apply electronically through the Queen’s online application portal at: https://dap.qub.ac.uk/portal/
+ Contact Details
|Supervisor Name:||Dr. Georgios Karakonstantis|
+44 (0)28 9097 6550