DC Field | Value | Language |
---|---|---|
dc.contributor.advisor | Park, KyoungSoo | - |
dc.contributor.advisor | 박경수 | - |
dc.contributor.author | Go, Younghwan | - |
dc.date.accessioned | 2019-08-25T02:45:06Z | - |
dc.date.available | 2019-08-25T02:45:06Z | - |
dc.date.issued | 2018 | - |
dc.identifier.uri | http://library.kaist.ac.kr/search/detail/view.do?bibCtrlNo=734389&flag=dissertation | en_US |
dc.identifier.uri | http://hdl.handle.net/10203/265190 | - |
dc.description | 학위논문(박사) - 한국과학기술원 : 전기및전자공학부, 2018.2,[viii, 92 p. :] | - |
dc.description.abstract | The explosive popularity of network communicating devices such as PCs, smartphones, tablets and IoT devices has driven massive growth in network data usage around the world. To meet the increasing network communication demand, a large number of networked systems are being constantly designed and deployed. Unfortunately, despite numerous efforts, we find that many networked systems often expose serious design problems in security, reliability and performance, which result in network attack, data corruption and perfor-mance bottleneck. This dissertation aims to identify the fundamental causes of these suboptimal designs and share our lessons in addressing them in real networked systems. As the first lesson in designing a secure networked system, we learn that a na?ve implementation of a complex policy often exposes security vulnerabilities, which leads to serious network attacks. When accounting for cellular data usage, we find that cellular ISPs either charge for TCP retransmission packets without robust retransmission detection or exclude them from billing by identifying retransmissions purely based on TCP header information. In either case, an attacker can easily exploit fake TCP retransmissions to either overcharge a targeted user or evade charging to use cellular data for free. To prevent such vulnerabilities, we develop a robust cellular data accounting system named Abacus, which accurately identifies fake TCP retransmission packets via efficient DPI. Abacus samples a few bytes strategically chosen from a retransmitted packet, unknown to the attacker, and efficiently compares against original packet payload for verification. As our second lesson, designing a reliable networked system often requires a clear understanding of users’ data usage patterns as a single corner case can result in catastrophic behavior that breaks data reliability. Mobile data clouds provide services to store and update user data with a guarantee for reliability and consistency. One important factor in mobile data synchronization is that these services operate over both structured and unstructured data such as a photo image linked with its metadata. However, we identify that existing services do not consider this close linkage and often fail to restore data to a consistent state after network or hard-ware failures, which are common for mobile users. As a result, unsynchronized mobile data can remain in a corrupted state and become unusable. We address this problem by presenting Simba, a novel mobile data synchronization service that ensures data consistency between tabular and object data through a logically unified table. To handle recovery after data corruption, we study mobile users’ data usage patterns and employ multiple local flags that can identify current state of the data after various failures. As our final lesson, we learn that building a networked system with high performance needs to take architectural differences into consideration as a reuse of existing designs from older platforms may result in an underperforming system. As a promising platform for high performance network applications, modern GPUs have been experimented with by many researchers to accelerate packet processing. Unfortunately, despite past research works that report performance improvement with GPUs, the PCIe data transfer delay, which is typically longer than individual packet processing time in GPU kernel, limits performance benefit in GPU-accelerated networked systems. As an alternate platform, an integrated GPU that shares DRAM with CPU can eliminate this overhead. However, we find that simply implementing the discrete GPU-based design on top of integrated GPU actually further degrades the performance. To avoid the problem, we redesign a networked system named APUNet, which addresses integrated GPU-specific issues such as memory contention and data synchronization overheads. We find that APUNet fully exploits the capacity of integrated GPU and operates in multi-10 Gbps networks with low cost and power. With the research works presented by Abacus, Simba and APUNet, we demonstrate that building a secure and reliable networked system with high performance requires design consideration on i) the security implications from any policy decision, ii) clear understanding of data usage patterns, and iii) architectural differences with existing design. While this dissertation focuses on addressing each design issue independently, we believe that our solutions are general and can be applied to other networked systems as well. We hope that our lessons and solutions serve as a stepping stone for building a next-generation networked systems platform. | - |
dc.language | eng | - |
dc.publisher | 한국과학기술원 | - |
dc.subject | networked system▼amobile networks▼acellular accounting▼amobile application▼amobile cloud▼aGPU acceleration▼aAPU▼aintegrated GPU | - |
dc.subject | 네트워크 시스템▼a모바일 네트워크▼a셀률러 과금▼a모바일 어플리케이션▼a모바일 클라우드▼aGPU 가속화▼aAPU▼a내장 GPU | - |
dc.title | Design consideration for secure and reliable networked systems with high performance | - |
dc.title.alternative | 안전하고 믿을 수 있는 고성능 네트워크 시스템 설계에 대한 고찰 | - |
dc.type | Thesis(Ph.D) | - |
dc.identifier.CNRN | 325007 | - |
dc.description.department | 한국과학기술원 :전기및전자공학부, | - |
dc.contributor.alternativeauthor | 고영환 | - |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.