CS262A Reading Summary 15

Concurrency Control Performance Modeling: Alternatives and Implications

r. Agrawal et al
Summary by Feng Zhou
10/8/2002

1. A single, physically justifiable database system model is built to facilitate performance studies of different concurrency control. The performance experiments show interesting performance results, the authurs give insightful comments and draw valid conclusions.
2. One important specific conclusion is: locking based cc schemes are generally more desirable with tight resource constraints, while optimistic cc schemes performs better with a lot of resource, high workload and high multi-programming level. This is understandable, because if restarts resulted from deadlocks are rare, locking based methods have smaller overhead. But under high workload, a lot of blocking makes concurrency too low, resulting low throughput.
3. Another important conclusion is that multiprogramming level in database systems should be carefully controlled. This is important and reasonable. For locking-based methods, when mpl is too high, many xacts block each others and deadlocks become common, thus overall throughput becomes very low. For optimistic cc methods, too much concurrent xacts result in too many verification failures, and thus restarts, also resulting in low throughput.

2 flaws:

One of the model assumption: concurrency control costs are neglectable, is not true in some cases. Examples include: dead-lock detection is often expensive and not neglectable; locking is often very expensive in parallel/distributed database environment. Adding these into the model may make it more general.

Another flaw is that only one parameter, Multi-Programming-Level, is varied in the experiments to get results. Other parameters, like number of terminals and database size, can possibly be changed to yield interesting results.