Implementation and experimental evaluation of the layered data replication algorithm

Abstract

In this thesis, we provide an implementation and experimental evaluation of the LDR algorithm [13]. LDR has the potential of being the basis of a distributed file system that achieves the three main goals that any File Distributed System wish to achieve, which are replication, performance and consistency guaranties and all these without using any high level functions such as distributed locking [18], embed physical writes to the data within a logical read [1, 19] or group communication [21], that negatively affect performance [22]. Our implementation can be used both in LAN and WAN, giving writers the opportunity to write all common types of files, including .doc, .pdf, .txt as well as .jpg images, and readers the certificate that they read the most up-to-date version of a file. The experimental evaluation took place on the Planetlab platform, and demonstrated that for up-to 250KB file size and a concurrent use from 10 writers and 20 readers, the system manages to serve all the clients and keeps its performance in high levels, using messages of 944 bytes as a default block size. Our expectations were fulfilled. As the results showed the implementation works for unbounded number of clients without violating the consistency guaranties and by having satisfying execution performance for large data objects replication.