Loading...
Please wait, while we are loading the content...
Similar Documents
Experimental Evaluation of Dynamic Data Allocation Strategies in a Distributed Database With Changing Workloads
| Content Provider | Semantic Scholar |
|---|---|
| Author | Carey, Marc B. Dewitt, David Frank, David Graefe, Goetz Richardson, Julian Shekita, Eugene J. |
| Copyright Year | 1995 |
| Abstract | Performance in distributed database systems is heavily dependent on the allocation of data among the sites of the database. The allocation of data is traditionally static and determined oo-line, using estimates of access frequencies. However, in many situations the access frequencies from varies sites in the database are not known a priori or uctuate with time thereby creating a changing workload. This paper showed that for a database with a changing workload dynamic reallocation of the data can signiicantly improve performance. A simple counter algorithm was presented which monitors the access frequencies in the system and moves the data so as to maximize the fraction of local accesses in the system. For our workload the algorithm ooered up to 30% performance gain over static allocations in a Local Area Network. We would expect even higher performance gains in a Wide Area Network as the cost of going across the network increases. Our experiments also showed that for certain workloads load balance must be considered when re-allocating the data. We presented a load sensitive counter algorithm which was shown to outperform the simple counter algorithm as well as static allocations for a class of workloads. While our experiments with a small-scale database make a practical case for dynamically re-allocating data in a changing environment, more work is needed to study various possible algorithms for dynamic data allocation and to test these algorithms on large-scale distributed databases. Several structural issues seem worthy of investigation such as appropriate block sizes for partitioning and the granularity at which statistics should be gathered for blocks in diierent relations. Our future plans also include lending theoretical support to our experimental results through analytical models. Acknowledgements We would like to thank the Exodus project for making Exodus available to the database community. We would also like to acknowledge the work of Brian Dewey in initially implementing the TPC-B workload running on top of the Exodus storage manager. |
| File Format | PDF HTM / HTML |
| Alternate Webpage(s) | http://members.es.tripod.de/jrodr35/ckim.ps |
| Language | English |
| Access Restriction | Open |
| Subject Keyword | Algorithm Allocation Brian Dewey Decimal Classification Distributed database Dynamic data EXodus Estimated Experiment IBM Tivoli Storage Productivity Center Load balancing (computing) Local Area Networks Workload |
| Content Type | Text |
| Resource Type | Article |
