Scale

Processing and data movement requirements of our environment cannot be met easily by monolithic systems. They have several major disadvantages:

• Partial upgrades are difficult or even impossible.
• Connectivity is limited by number of ports and memory bandwidth.
• Larger machines are more expensive to build.
• Error in a part of the system causes failure of the whole system.
• Maintenance causes unavailability of the whole system.

These reasons drive the deployment of clusters built from commodity computing components. They have the numerous advantages (partly from [McC99,FGC$^$97]):

• Per unit cost of the system is smaller because of mass production.
• Incremental upgrade with state of the art commodity components is possible.
• Availability is high, as single failures can be tolerated.
• Performance of the sum of units is generally higher than the performance of a single large machine. This stems from the fact that resources do not have to be shared.
• Clusters of independent units are well suited for applications with simultaneous independent users or applications.
• Each application can comprise richer functionality, as the whole node is exclusively available for processing.
• Due to their openness, computing clusters are easier to adapt to specific needs.

Spreading the workload over a large number of processors and network ports helps to overcome the limitations of processing power in single machines. A system that allows such graceful extension in order to meet its performance specification is said to be scalable. This quality becomes more and more important for the software part as it is difficult to manage large quantities of independent units. Furthermore multiplicative growth of messages in our system limits scalability. Each node has to keep state information about all other communication participants. It has been shown that at a number of about 700 communication partners and 300 threads to handle this load, 70% of processing is spent in the operating system kernel[FGC$^$97, pp. 86-87]. This exhibits one of the problems with we have to deal with when it comes to design of communication middleware.

Requirement 2 (Scalability)   Software for construction of a distributed data acquisition system shall support scalability of the cluster. Scalability means, that the increase of overhead with additional participants in the system is predictable (at best linear) and that the other performance requirements can be met.