Management of Multilevel, Multiclient Cache Hierarchies with Application Hints
ACM Transactions on Computer Systems (TOCS)
| Hi-index | 0.01 |
The gap between processor speeds and speed of technologies providing data is increasing. This causes the performance of client applications to be limited by the performance of storage devices, networks and buses. Furthermore, the number of computers that share these data access resources is growing exponentially. Caching, prefetching and parallelism are some of the techniques used today to cope with data access latency and system scalability to support more users. However, there is a lack of support for caching, which is widely and freely available over the Internet. This thesis makes several contributions to the research area of high-performance distributed data access. The first contribution is the design of a globally-distributed Internet cache service, called Storage Embedded Networks (SEN). SEN improves user response times and data service scalability of the Internet by ubiquitous caching. It is composed of routers containing storage that caches some of the objects that are pass through via snooping. Requests are checked every hop, ensuring transmission of the closest copy on the data path and upstream load reduction. SEN architecture also addresses the flash crowd problem. The second contribution, called Adaptive Caching using Multiple Experts (ACME), is an automated caching scheme that combines the expertise of multiple replacement algorithms to manage the SEN caches. ACME uses machine learning techniques to select the best current replacement policy. It improves the hit rates over static caching policies leading to lower client response times and more network savings. Each cache node can adapt itself to the application workload it observes. ACME eliminates manual tuning from multi-level caches.