CS 736 - Spring 2007 - Paper Discussion

Summary:
THe paper describes the overall design of an architecture that uses lightweight disks instead of full-blown servers to service requests for data.

Problem:
When increasing the number of clients in traditional network file systems, the load of the servers grow to the point where the servers become the bottleneck. The authors of this paper were trying to improve the scalability.

Contributions:
* An argument of the potential cost benefits of dropping the general purpose computer part of the actual data servers in favor of their lightweight NASD

* Discussion of some of the implementation difficulties and possible design decisions, like how security should be implemented in the system

* Tests to show the potential performance benefits of NASD, by comparison to existing NFS implementations.

Flaws:
They only evaluated up to 10 clients. This seems like a pretty low number, especially considering that there is a noticable increase in the slope of the line right at 10 clients.

Performance:
As was mentioned before, the authors were trying to improve the scale out ability with respect to the number of clients. They were attempting to create a different architecture where most of the "servers" were unusually lightweight disks.

Summary

In this paper, the authors� present the architecture and prototype implementation of Network-Attached Secure Disk for storage. By removing file-servers from the path between clients and peripheral networks and by transferring data directly to the clients, NASD results in scalable storage bandwidth by embedding disk management functions into the device.
Due to significant increase in transistor density in in-expensive disk technology, sophisticated special purpose functional units can be integrated on the disk. This results in lower cost and an increase in disk performance.

Problem Description

Distributed file systems are accessed through a file-server which implements all the disk management and cryptography functions. With the increase in the number of clients, the pressure on these servers increase as a result, the bandwidth does not scale well. In this paper, the authors exploit the fact that the transistor density on an ASIC has increased significantly as a result disk management and cryptography functions can be integrated inside a disk and the file-servers are avoided altogether. As a result, the bandwidth scales quite well with an increase in the number of clients.

Summary of Contributions

The major contribution of this paper is to propose the NASD architecture that promises a scalable increase in bandwidth as the number of clients increase. This is achieved by transferring data directly to the clients. Due to an increase in the transistor density, disk management and cryptography functions can be integrated inside the disk. This results in a scalable bandwidth by avoiding the file-servers from the data transmission path.
The authors also suggest using Active Disks for further increase in disk performance. Incase of active disks, on-drive computations are exploited by providing full application level programmability of drives. As a result, code to manage data stored in the disk is executed near the data before it is placed on the interconnect network. This provides the ability to customize functionality for specific data-intensive applications.

Flaws

Some of the flaws in this paper are as follows.
1. The authors present the architecture of the NASD at a very high level. They do not explain different components of NASD in details. As a result, the graphs presented seem to make very little sense. This is due to the fact that it is not explained what major features of NASD result in significantly scalable bandwidth.
2. NASD uses File Manager for access control and namespace consistency. The overhead incurred because of the File Manager is not discussed in the paper.

Performance

The paper mainly focuses on proposing an architecture that results in scalable bandwidth for a distributed file system. To achieve this goal, the authors propose NASD which transfers data directly to the clients and avoids the bottle presented by a file-server.