Clustera: A Data-Centric Approach to Scalable Cluster Management

                         David J. DeWitt
                 Microsoft Jim Gray Systems Lab &
 Computer Sciences Department, University of Wisconsin at Madison

Twenty-five years ago, when we built our first cluster management
system using a collection of twenty VAX 11/750 computers, the idea of
a compute cluster was an exotic concept. Today, clusters of 1,000
nodes are common and some of the biggest have in excess of 10,000
nodes. Such clusters are simply awash in data about machines, users,
jobs, and files. Many of the tasks that such systems are asked to
perform are very similar to database transactions. For example, the
system must accept jobs from users and send them off to be executed.
The system should not "drop" jobs or lose files due to hardware or
software failures. The software must also allow users to stop failed
computations or "change their mind" and retract thousands of submitted
but not yet completed jobs. Amazingly, no cluster management system
that we are aware of uses a database system for managing its data.

In this talk I will describe Clustera, a new cluster management system
we have been working for the last three years. As one would expect
from some database types, Clustera uses a relational DBMS to store all
its operational data including information about jobs, users, machines,
and files (executable, input, and output). One unique aspect of the
Clustera design is its use of an application server (JBoss currently)
in front of the relational DBMS. Application servers have a number of
appealing capabilities. First, they can handle 10s of 1000s of
clients. Second, they provide fault tolerance and scalability by
running on multiple server nodes. Third, they multiplex connections to
the database system to a level that the database system can
comfortably support. Compute nodes in a Clustera cluster appear as web
clients to the application server and make SOAP calls to submit
requests for jobs to execute and to update status information that is
stored in the relational database.

Extensibility is a second key goal of the Clustera project.
Traditional cluster management systems such as Condor were targeted
toward long-running, computational intensive jobs. Newer systems such
as Map-Reduce are targeted toward a specific type of data intensive
parallel computation. Parallel SQL database systems represent a third
type of cluster management system. The Clustera framework was designed
to handle each of these classes of jobs in a common execution and data
framework.