Tiwana et al (2010) Location, Location, Location! Modeling Data Proximity in the Cloud (HotNets)

In a cloud computing environment, the response time of an application is correlated to the computation node’s proximity to the data.

According to the paper, the current practice is to use the closest server to handle a user’s job. Then, by predicting the future access patterns based on history, move the data to closer to the computation nodes.

This paper do the reverse: Expose the information of the location of the data to the application, and let the application to decide where to launch the computation. Different storage technology, however, has different replication topologies. A replication topology is a term in the paper to define how a read/write action involves different nodes. For example, if the storage is mirrored, the read involves any copy of the mirror but write involves sending the write instruction to all mirrors. These actions can be formulated as a directed acyclic graph (DAG). The paper suggests that, to have a middle layer called Contour between the storage system and the application. The storage system can return the replication topology of a particular file when Contour requests. This topology includes the exact location of nodes involved. Contour system provides function to the application so that it can tell which are the nodes that an application should launch the computation so that the computation is closest to the data, or so that the computation can satisfy certain latency constraints.

Bibliographic data

@inproceedings{
   title = "Location, Location, Location! Modeling Data Proximity in the Cloud",
   author = "Birjodh Tiwana and Mahesh Balakrishnan and Marcos K. Aguilera and Hitesh Ballani and Z. Morley Mao",
   booktitle = "Proc HotNets",
   year = "2010",
}