Research Challenges in Flash Memories
Intervenant(s) : Sivan Toledo (Tel-Aviv University and MIT)
Flash memories (the non-volatile memory chips in USB memory sticks, memory cards for cameras, some music players, etc.) pose some interesting research challenges in algorithms, systems, and architecture. The challenges arise because flash is weird. You must erase large blocks before you can overwrite data in them (reads and writes have smaller granularity than erasures). Writes are slower than reads. Memory cells wear out after a certain number of writes (only 10,000 in new devices and probably going down). These challenges are non-trivial even in the traditional applications of flash (e.g., USB sticks, digital cameras, small embedded systems). But they become even more interesting as flash is on the cusp of large-scale deployment in laptops, servers, and pretty much any computer system that we now use. The talk will cover the basics of flash memories and the challenges that flash poses in algorithms, systems, and architecture. I will also describe some theoretical results on the effectiveness of wear leveling and erasures, on several flash-based storage subsystems that we have built, and on the measured performance characteristics of flash-based devices (USB, disk replacements, and memory cards).