Improving Proxy Cache Performance: Analysis of Three Replacement Policies Dilley, J.; Arlitt, M. A journal paper of IEEE Internet Computing, Volume: 3 Issue: 6, Nov.-Dec Page(s):

Outline Introduction Related work New replacement policies Simulation Conclusion

Introduction The goal of web cache  Reducing the bandwidth demand on the external network.  Reducing the average time it takes for a web page to load.

The location of web cache  Client side  Proxy side  Web server side

Cache replacement policy The reason why we need replacement  A cache server has a fixed amount of storage.  When this storage space fills, the cache must choose a set of objects to evict.

Cache replacement policy (cont.) The goal of cache replacement policy  To make the best use of available resources, including disk and memory space and network bandwidth.

Web proxy workload characterization The characteristic of workload that we interest  The patterns in the number of objects referenced and the relationship among accesses.

Web proxy workload characterization (cont.) The workload that we choose to simulate  Traces of actual live execution The drawback of traces of actual live execution  It does not capture changing behavior or the behavior of a different user set.

Measures of efficiency  Hit rate It’s the most popular measure.  Byte hit rate It’s interesting on the limited bandwidth and expensive resource.  CPU and I/O system utilization  Object retrieval latency (or page load time)

An optimal cache replacement policy  To know a document’s future popularity and choose the most advantageous way to use its finite space.

The heuristics to approximate optimal cache replacement  Caching a few documents with high download latency might actually reduce average latency.  To maximize object hit rate, it is better to keep many small popular objects.

Related work Replacement algorithms  Least Recently Used (LRU) It works well when there is a high temporal locality in workload. It’s data structure is doubly linked list to implement.

Replacement algorithms (cont.)  Least Frequently Used (LFU) It evicts the the objects with the lowest reference count. It’s data structure is heap. It’s drawback is the problem of cache pollution.

Replacement algorithms (cont.)  LFU-Aging (LFUA) It considers both an object’s access frequency and its age in cache. It avoid the problem of cache pollution.

Replacement algorithms (cont.)  Greedy Dual-Size (GDS) It takes size and a cost function for retrieving objects into account.

New replacement policies LFU with Dynamic Aging (LFUDA)  It adds a cache age factor to the reference count, when a new object is added or an existing object is referenced.

New replacement policies (cont.) GDS-Frequency (GDSF)  It considers reference frequency in addition to object size.  It assigns a key to each object.  Key= (reference count of object / size ) plus cache age

Simulation Two types of cache replacement algorithm  Heap-based such as GDSF, LFUDA, LRU_H It’s data structure is heap. Let memory usage grow to a high watermark before examining any cache metadata

Two types of cache replacement algorithm (cont.)  Listed-based such as LRU_L It’s data structure is linked list Examining object metadata more often to implement the LRU reference age.

The result of simulation Hit rate

Byte hit rate

CPU utilization

Finding LRU_L has higher and more variable response times than LRU_H because of the extra time spent computing the reference age and choose the object to evict. LRU_L ’ s extra time seems to translate into a higher hit rate for LRU_L. GDSF shows consistent improvement over LRU.

Conclusion More sophisticated policy (heap-base for example) can be implement without degrading the cache ’ s performance. The bottleneck is on I/O (network and disk), not on CPU utilization. To achieve the greatest latency reduction, to optimize the cache hit rate.

While the bandwidth id dear, we must focus on improving the byte hit rate. The impact of consistency validation of cached objects becomes more significant.