100 likes | 204 Vues
This paper presents a design for a database cluster architecture tailored for dynamic content web applications, evaluated using the TPC-W benchmark. The study illustrates that TPC-W scales effectively with increased cluster size, with conflict-aware scheduling demonstrating the most significant positive impact on performance. Additionally, the paper discusses the secondary effects of load balancing and finds that optimizing for locality does not improve performance. The findings contribute to ongoing discussions about efficient scheduling, caching, and replication strategies for dynamic content delivery.
E N D
Written by C. Amza, A. L. Cox, W. Zwaenepoel A Comparative Evaluation of Transparent Scaling Techniques for Dynamic Content Servers Presented by Chen Zhang2006-10-04
Overview • What does the paper do? • Designed a DB cluster architecture oriented at dynamic content web sites • Evaluated transparentscaling technique combinations against TPC-W • Experimental Findings • TPC-W scales well with cluster size increase • Conflict-aware scheduling with most beneficial impact • Load balancing has secondary impact • Optimizing for locality has no impact
Who is who • Sch • Scheduling • Caching • Seq • Adding sequence number • DP • Query queuing for ordering • Admission Control • DB Consistency • 1-copy-serializability
Scheduler and DP • Scheduler • Parse individual queries – context-aware • Backlog for all replicated ops for active trans. • Current load of DB • Consistency among other schedulers • DP • Conflict-aware • Lock tables, not passed to DB • In order • Out-of-order • Query prioritizing (a little bit of scheduling) • Admission Control
Scheduling and load balancing • Scheduling • Synchronous Replication • Content-Aware asynchronous replication • Conflict-aware deadlock avoidance • At DP, “lock tables” • Load balancing • Generic: Round Robin, Shortest Queue First • Content-aware • Shortest Execution Length First (SELF) • Locality-aware Request Distribution (LARD)
Caching • Happen at schedulers • Function • Cached query results for all reads • Forwards “lock tables” • Constraint • Require low write frequency • Size • LRU as Replacement strategy • Consistency between different scheduler cache.
Experiment • TPC-W Benchmark • Cluster • max 8 DB machines • 2 schedulers • Software • Apache • PHP • Mysql • Tested Combinations • Base • BestSync • ConflA • ConflACache
Discussion • Recall Ganymed • Scaling • Cache • Dynamic content replication • About scheduler and DP • About degree of consistency • ……