The Design Of Instruction Analysis Based Non-volatile Processor System

Nowadays,with the rapid development of embedded technology,the Internet of Things(IoT)is gradually influence people's lives and plays an important role.In order to meet mobility and profitability,most of these wearable devices are powered by batteries.However,the shortcomings of battery like size limitation,safety consideration and recharge inconvenience limit the application scenarios of IoT devices.Therefore,energy harvesting system has been widely investigated as a promising alternative as power supply for IoT devices.Energy harvesting system can generate electric energy from ambient energy sources,such as solar and wind energy.However,the unstable ambient energy source is an intrinsic challenge of energy harvesting system.All of them are unstable.With unstable power supply,the processor may be interrupted frequently,where traditional CMOS-based volatile processors will lose intermediate data upon power failures.After power resumes,the system will be rebooted.Frequent turning-off and rebooting will place extra burden on limited power budgetTo address this problem,non-volatile processor(NVP)has been proposed,which attaches non-volatile memory to the volatile logics.Due to the physical characteristics,non-volatile memory(N VM)can hold data even when power is off.Thus,when a power failure occurs,program state information can be backed up,and be recovered after power resuming for program continuation.The prime objective of NVP design is to guarantee the system correctness with backup/resumptions.On the basis of that,optimizations can be conducted to reduce the cost of backup and resumption,in order to improve system performance and energy utilization.However,there are related work focusing on a part of NVP system such as cache or main memory levelIn this paper,we target an online backup manager which guides the backup to achieve the maximal forward progress on system level perspective.To improve forward progress to the most extent,we choose the backup location by iteratively investigating the feasibility of executing next instruction,in which way the program can continue execution until the furthest position before successful backup.Furthermore,due to the limitation of tape out,we develop a more completed simulator based on gem5 for NVP system,which supports more flexible memory hierarchies and also equips with accurate system-level energy models.Finally,experimental results show that compared to the instant backup method,the proposed policy achieves 24.2%forward progress improvement on average.