Efficient Management Techniques for Improving the Performance and Optimizing Power Consumption in Smartphone System

Abstract

In spite of many years' efforts of smartphone development, the needs for high performance and low energy smartphone are still increasing. However, the current mobile platforms in smartphone systems are built on top of the traditional desktop or server operating systems like Linux. Traditional operating systems have different design goals and workloads from the smartphone systems. For example, the power consumption is not a major issue for traditional operating systems since machines are always plugged into a power source. CPU is a major source of power consumption on desktop machines where the default power management policy is that the CPU stays on or runs at a high frequency [90]. In contrast, the most important issue for smartphone systems is an efficient power saving management policy as the mobile devices have limited battery life. As for the hardware I/O components, traditional components like CPU, memory and storage are found in desktop or laptop machines but location based service using GPS, camera or various built-in sensors were not part of the traditional desktop machines. These new components and new system requirements of the smartphone systems dictate the major redesign of the traditional operating systems. In this thesis, we analyze the characteristics of smartphone workloads based on real experiment results and show that smartphone applications have very different file access patterns than the traditional operating system workloads. In the traditional desktop systems, read access dominates in the file access of applications. But in the smartphone applications, write access is more intensive than read access. In particular, write-once file access is much more frequent than others in smartphone workloads. Different workload characteristics require different system design. We analyze the write-once file access pattern in the smartphone and show that most write-once file access pattern is due to the frequent synchronous call performed by the SQLite journaling system, which tries to guarantee the data consistency between DRAM and Flash storage. Such frequent writing to storage by SQLite journaling system consumes a lot of power, which was not an important design issue in traditional operating system design. Previous researches show that up to 40% of the total system power is consumed by the memory/storage subsystem [93]. Frequent synchronous calls (e.g., fsync) to storage not only consume power but also degrade the performance in mobile systems. Recent progress in emerging non-volatile memory technologies such as Phase Change Memory (PCM) or spin-transfer torque random-access memory (STT-RAM) can potentially alleviate this problem. In this thesis, we first propose a new PCM based journaling scheme, which provides high storage performance without any loss of reliability. One important part of the journaling scheme is the buffer cache system. We propose a new buffer cache management scheme which classifies write-only-once data and aggressively evicts them from the cache to improve cache space utilization by exploiting write-once file access pattern characteristics in smartphone workloads. This new buffer cache scheme improves file access performance by maintaining hot data in memory thereby servicing subsequent requests without storage accesses. We implemented this new journaling scheme on the Linux mobile system. Performance study shows that the proposed buffer cache scheme improves the cache hit ratio by 5-33% compared to the conventional systems. The power consumption of memory is also improved by 64% on average. Storage performance is improved by 20-278% and the power consumption is reduced 63% on average compared to the conventional journaling system. The execution times of real applications such as games, video players, and web browsers are improved by 23-69%. Our research demonstrates that the new journaling system that introduces the non-volatile PCM can greatly improve storage performance at a little extra hardware cost and reduce power consumption compared to the original journaling system without any loss of reliability. Different manufacturing processes often require the need for different designs of the systems. To enhance the quality of the user experience, modern smartphones have various hardware components like GPS and sensors. We investigate location based service - another important workload that dictates different system design for smartphone than the conventional desktop or server operating system. In order to calculate the exact position of the smartphone based on the signals from the satellite, the smartphone systems turn on the GPS receiver most of the time. This prevents the system from going to sleep when no other activities are going on in the system. To minimize the power consumption associated with GPS, we present a new location estimation scheme called Adaptive Location Detection (ALD). The performance evaluation shows that ALD reduces the GPS power consumption by 49.5% on average, while satisfying the accuracy requirement of the most smartphone applications.;최근 스마트폰에 대한 높은 인기와 활발한 연구에도 불구하고, 여전히 고성능과 저전력에 대한 요구는 계속해서 늘어나고 있다. 스마트폰의 운영체제의 설계에 있어서 성능과 에너지 효율성은 매우 중요한 고려사항이다. 하지만 현재 널리 사용되고 있는 스마트폰 모바일 플랫폼은 전통적인 데스크탑 운영체제에 기반한 구조로 플래시에 집중되어 설계되어있어 스마트폰 고유의 특성을 반영하지 못하고 있다. 이로 인해 스마트폰의 주기적인 스토리지 동기화로 인한 성능 저하가 발생하고, 읽기보다 쓰기가 빈번하게 발생하는 스마트폰 애플리케이션의 특성을 전혀 고려하지 않은 버퍼 캐시 설계로 인해 캐시 공간의 비효율성이 발생한다. 또한, 다양한 센서가 내장되어 있고 무선 네트워크가 가능한 모바일 기기라는 스마트폰의 특성상 이러한 요소들이 발생시키는 전력 소모를 최소화하여 배터리 수명을 최대한 늘릴 수 있는 새로운 기법의 설계가 필요하다. 본 학위 논문에서는 스마트폰 고유의 특성을 분석하고 스마트폰의 특성을 반영하는 새로운 시스템 아키텍쳐와 다양한 기법들을 제안하여 스마트폰의 성능을 향상시키고 에너지 효율성을 높일 수 있게 한다. SQLite가 발생시키는 동기적 쓰기가 스마트폰의 스토리지 성능에 끼치는 영향을 분석하기 위해, 동일한 데이터 쿼리에 대해 저널링을 수행할 때와 그렇지 않을 때의 성능을 정량적으로 측정하였다. 이러한 분석결과를 바탕으로 신뢰성을 희생시키지 않고 기존 저널링 방식의 동기적 쓰기 오버헤드를 줄이기 위해 비휘발성 램을 활용한 새로운 아키텍쳐를 제안한다. 또한 스마트폰 애플리케이션의 파일 접근 중 상당 부분이 1회성 쓰기 참조라는 것을 분석하고 이러한 특성을 고려한 버퍼캐쉬 관리 기법을 제안한다. 스마트폰 센서 중 상당한 전력을 소모하는 GPS의 문제를 해결하기 위해 셀 기반의 위치 추정방식을 사용하여 스마트폰 애플리케이션이 요구하는 정확성에 맞게 적절한 위치정보를 제공하면서도 전력을 최대한 줄일 수 있는 새로운 위치 추정 기법을 제안한다. 각각의 연구 주제에 대해 본 논문에서는 다양한 벤치마크 프로그램, 리눅스 기반의 구현과 모의 실험을 통해 새롭게 제안한 기법의 성능을 검증한다.