Improving DRAM Performance, Security, and Reliability by Understanding and Exploiting DRAM Timing Parameter Margins

Abstract

Characterization of real DRAM devices has enabled findings in DRAM device properties, which has led to proposals that significantly improve overall system performance by reducing DRAM access latency and power consumption. In addition to improving system performance, a deeper understanding of DRAM technology via characterization can also improve device reliability and security. These can beseen with the recent discoveries of 1) DRAM-based true random number generators (TRNGs), a method for generating true random numbers using DRAM devices which can be used in many applications, 2) DRAM-based physical unclonable functions (PUFs), a method for generating unique device-dependent keys for identification and authentication, and 3) the RowHammer vulnerability, a phenomenon where repeatedly accessing a DRAM row can cause failures in unaccessed neighboring DRAM rows.To advance DRAM-based discoveries and mechanisms, this dissertation rigorously characterizes many modern commodity DRAM devices and shows that by exploitingDRAM access timing margins within manufacturer-recommended DRAM timing specifications, we can significantly improve system performance, reduce power consumption, and improve device reliability and security. First, we characterize DRAM timing parameter margins and find that certain regions of DRAM can be accessed faster than other regions due to DRAM cell process manufacturing variation. We exploit this by enabling variable access times depending on the DRAM cells being accessed, whichnot only improves overall system performance, but also decreases power consumption. Second, we find that we can uniquely identify DRAM devices by the locations offailures that result when we access DRAM with timing parameters reduced below specification values. Because we induce these failures with DRAM accesses, we cangenerate these unique identifiers significantly more quickly than prior work. Third, we propose a random number generator that is based on our observation that timingfailures in certain DRAM cells are randomly induced and can thus be repeatedly polled to very quickly generate true random values. Finally, we characterize the RowHammersecurity vulnerability on a wide range of modern DRAM chips while violating the DRAM refresh requirement in order to directly characterize the underlying DRAM technology without the interference of refresh commands. We demonstrate with our characterization of real chips, that existing RowHammer mitigation mechanisms eitherare not scalable or suffer from prohibitively large performance overheads in projected future devices and it is critical to research more effective solutions to RowHammer.Overall, our studies build a new understanding of modern DRAM devices to improve computing system performance, reliability and security all at the same time.