Constructing enhanced file system and memory abstractions on public cloud

MPhil Thesis Defence


Title: "Constructing enhanced file system and memory abstractions on public cloud"

By

Mr. Zhonghua Sheng


Abstract

With the development of cloud-based systems and applications, a number of major 
technical firms have started to provide public cloud services. A challenge 
behind the cloud computing is how to store, share and process the users’ data 
securely and efficiently. Therefore, we need to enhance the file system and 
memory abstractions for the cloud computing paradigm.

On the traditional computing architecture, the user have full control over 
their data on their computer or in their private datacenter. However, when 
users store private data in shared datacenters, they lose control over how the 
data are stored and accessed. Multiple classes of personnel may access the 
physical storage media and potentially read the data. While strong 
cryptographic methods can protect user files from unauthorized accesses, they 
incur computational overhead, and make it difficult for the infrastructure 
provider to optimize the storage space with effective compression and 
deduplication. To provide strong protection on user data, we design a new file 
system called BIFS (Bit-Interleaving File System). Focusing on the privacy 
protection of the on-disk state, BIFS re-orders data in user files at the bit 
level, and stores bit slices at distributed locations in the storage system. 
While providing strong privacy protection, BIFS still retains part of the 
regularity in user data, and thus enables the infrastructure provider to 
perform a certain level of space optimization (e.g., compression).

Storing the users’ data with strong privacy protection is not enough. Sharing 
and processing the user’s data efficiently on the cloud infrastructure remains 
a factor which prevents many users from stepping into the cloud computing era. 
Message passing introduces too much burden to the users for programming. 
Traditional Distributed Shard Memory (DSM) design fails in scalability and 
efficiency. Focusing on the efficiency and flexibility, while changes the 
user’s programming behavior as less as possible, we combine the advantages of 
transactional memory and distributed shared memory to design a scalable 
distributed transactional memory.

We implement BIFS and SDTM to examine their performance characteristics. The 
cloud storage service used for our implementation of BIFS is the Amazon Simple 
Storage Service (S3). The comparison with several existing network or 
Internet-based file systems shows that BIFS provides robust file system 
functions with satisfactory throughput on S3. The evaluation of SDTM also shows 
it is both scalable and efficient.


Date:			Wednesday, 16 May 2012

Time:			10:30am – 12:30pm

Venue:			Room 3501
 			Lifts 25/26

Committee Members:	Dr. Lin Gu (Supervisor)
 			Dr. Jogesh Muppala (Chairperson)
 			Dr. Qiong Luo


**** ALL are Welcome ****