Research On High-performance And Low-power Mobile Computing Based On Non-volatile Memory

With the development of the Internet,the maturity of mobile communication technology and the popularity of distributed computing,mobile computing gradually comes into people's daily life.Mobile computing technology makes people can easily achieve information interaction and resource sharing anytime and anywhere.With more and more intelligent devices involved into mobile computing activities,in the foreseeable future,mobile computing will have a great development prospect and a broader development space,which will completely change people's lifestyle and habits.With the rapid development of mobile computing technology,the emerging application scenarios and requirements bring many new challenges.Performance and energy are two major problems which hinder the development of modern technologies.To deal with the increasingly complex tasks,mobile computing devices need to continuously improve the performance to guarantee tasks' normal execution.Considering the limited battery capacity,how to reduce the energy consumption as much as possible while ensuring tasks' completion is an urgent problem to be addressed.In addition,the mobile computing activities not only require flexible network management and control,but also need to ensure the message transmission with a lower latency,especially when the network scale is very large.The emerging non-volatile memory technologies bring a new opportunity of solving the above issues.Different from the traditional SRAM(Static Random Access Memory)and DRAM(Dynamic Random Access Memory).NVMs have the advantages of low power,non-volatility and high density.The low power feature can reduce the increasing energy consumption of memory system,non-volatile characteristics eliminates the performance loss caused by the data consistency issue and the high density improves the spatial expansibility which can better adapt to the scenes with large space requirements.However,compared with the traditional memories,NVMs have a lot of unavoidable problems,such as longer access delays,non-uniform read/write operations and limited write endurance.Therefore,how to use NVM techonologies to solve the problmes in current and future mobile computing while overcoming the inherent constraints is an important problem worthy to be studied.In order to eliminate the obstacles that hinder the development of high-performance and low-power mobile computing,in this paper,we study the design and optimization strategy of the emerging non-volatile memory system.This paper first studies the problem of increasing energy consumption of moble terminals and uses NVM's low power feature to explore corresponding solutions.Although allocating tasks to NVM can achieve energy savings,it will increase the execution time since NVM has lower performance compared with DRAM.Since many mobile computing tasks have real-time constrainsts,the increament of execution time will result in potential risks that some tasks cannot complete before their deadlines.To this end,in this paper,we adopt a hybrid NVM and DRAM main memory architecture to fully use the high performance of DRAM and the low power of NVM to achieve energy saving as much as possible while ensuring the task set keeps schedulable.For an aperiodic task set,we propose an Offline-Aperiodic scheduling algorithm.But for a periodic task set,the static priority scheduling strategy and the dynamic priority scheduling strategy are studied,respectively.We take two representative algorithms,RM(Rate Monotonic)algorithm and EDF(Earliest Deadline First)algorithm,as examples to present our Offline-RM and Offline-EDF scheduling algorithms.The proposed offline scheduling algorithms reserve the Worst-Case Execution Time(WCET)for ael tasks to guarantee the task set is schedulable.Experimental results show,compared with the traditional scheduling algorithm for DRAM-based main memory,the three proposed offline scheduling algorithms for hybrid main memory architecture achieve average energy saving of 28.8%,25.69%and 38%.respectively.Since a task's actual execution time is much less than its worst-case execution time,the slack time generated by the completed tasks can be further allocated to the unfinished tasks so that more tasks can execute in NVM.Therefore,based on the research of offline scheduling algorithms,we propose energy-aware online scheduling algorithms.For an aperiodic task set,when a task completes,we reclaim the slack time it generates and check if more tasks can be allocated to NVM.But for a periodic task set,the allocation of slack time may affect the execution of subsequent tasks since tasks arrive periodically.Moreover,the task preemption further increases the risk that the task set cannot keep schedulable.To this end,the proposed Online-RM algorithm focuses on the closest deadline to reclaim the slack time.But for the EDF scheduler,since it schedules tasks more flexibly,we manage the available slack time from the global perspective.The Online-EDF algorithm avoids insecure time allocations by introducing the data structure Ready-Queue and Preempt-Queue.In order to reduce the overheads caused by task migrations between NVM and DRAM,the proposed online algorithms for periodic tasks ensure each task instance can be migrated at most once.The experiments show the proposed online scheduling algorithms achieve further energy savings compared with the offline algorithms.In addition,to deal with the increasingly complex tasks,mobile computing devices need to continuously improve the performance to guarantee tasks' execution.However,the journaling filesystem,which is adopted in current mobile devices,hinders the improvement of system performance.With the value of data is becoming more and more important,in order to guarantee the reliability,journaling filesystems exploit the journaling technique to avoid the data inconsistence caused by unexpected system crach.Journaling technique first writes data to a journal area and then writes it to its home location.By using journaling techniques,the partially committed data can be recovered from the journal area.Obviously,the journaling mechanism causes a large number of extra writes operations.The reason of this problem is the DRAM-based Buffer Cache is volatile.When system crach occurs,the unsaved data will be lost.Different from DRAM,NVM is non-volatile.The modified data can be maintained in NVM before being written back to the filesystem,thus eliminating the extra I/O operations caused by journaling.Therefore,we adopt the hybrid NVM and DRAM buffer cache architecture to reduce the journaling overhead.In order to better utilize this novel architecture,we propose a Journaling-Aware Page Management(JAPM)policy which optimizes the journaling technique and overcomes the disadvantages of NVM in terms of write performance and lifetime.In addition,data in one transaction may be dispersed in different storing media.In order to guarantee the atomicity of the transactional execution in the hybrid cache architecture,we further propose a Partial In-Place Commit(PIPC)journaling scheme to coordinate the different committing patterns.The experimental results show the proposed mechanisms effectively reduce the journaling overhead and improve the system performance without any loss of data reliability.The implementation of high-performance and low-power mobile computing on one hand depends on the enhanced ability of mobile devices;on the other hand it relies on a flexible and efficient mobile network.In the future,we not only need to keep the flexibility and controllability of the large scale network,but also should reduce the control energy as much as possible and ensure the high-performance delivery of the packets.With more and more devices involved into mobile computing activites,current network architecture and communication equipments face new challenges in terms of controllability and forwarding performance.Therefore,in this paper,we present a software-defined moblile computing architecture and optimize the problem of high control overhead.Meanwhile,we study the complex mobility issue since the software-defined architecture cannot adapt to the frequent mobility environment well.The introduction of the software-defined architecture realizes the flexible control of the network,which further increases the burden of traditional communications facilities.Modern switches generally use TCAM(Ternary Content Addressable Memory)for packets matching,but the SRAM-based TCAM is far from satisfying the requirements of current and future mobile computing,both in terms of storage capacity and power conconsumption.Considering the high density and low leakage power of NVM technologies,we propose a hybrid TCAM architecture,which is composed of SRAM-based TCAM and NVM-based TCAM.To better utilize the hybrid architecture,we propose a Rule Scheduler for Hybrid TCAM(RSHT)algorithm.RSHT puts the most popular rules in nvTCAM(NVM-based TCAM)to improve the forwarding performance of the hybrid TCAM and a small part of rules in sTCAM(SRAM-based TCAM)to reduce the writes on NVM,thus improving the update efficiency.The experimental results show that the software-defined architecture effectively reduces the control overheads while maintaining the flexibility of the mobile computing network.Meanwhile,the proposed hybrid TCAM architecture significantly improves the forwarding performance of traditional communication facilities,thus providing reliable guarantees for the implementation of high-performance and low-power mobile computing.