A (Nearly) Free Lunch

NAND flash is a key storage technology in modern computing systems. Without it, many devices would probably not exist today or would at least not benefit from as many features. The very large success of this technology motivated massive efforts to scale it down in order to increase its density further. However, NAND flash is currently facing physical limitations that prevent it reaching smaller cell sizes without severely reducing its storage reliability and lifetime. Accordingly, in the present thesis we aim at relieving some constraints from device manufacturing by addressing flash irregularities at a higher level. For example, we acknowledge the fact that process variation plus other factors render some regions of a flash device more sensitive than others. This difference usually leads to sensitive regions exhausting their lifetime early, which then causes the device to become unusable, while the rest of the device is still healthy, yet not exploitable. Consequently, we propose to postpone this exhaustion point with new strategies that require minimal resources to be implemented and effectively extend flash devices lifetime. Sometimes, our strategies involve unconventional methods to access the flash that are not supported by specification document and, therefore, should not be used lightly. Hence, we also present thorough characterization experiments on actual NAND flash chips to validate these methods and model their effect on a flash device. Finally, we evaluate the performance of our methods by implementing a trace-driven flash device simulator and execute a large set of realistic disk traces. Overall, we exploit properties that are either neglected or not understood to propose methods that are nearly free to implement and systematically extend NAND flash lifetime. We are convinced that future NAND flash architectures will regularly bring radical physical changes, which will inevitably come together with a new set of physical properties to investigate and to exploit.