130 likes | 218 Views
Squeezing More Bits Out of HTTP Caches. Jeffrey C. Mogul Presented by: Zhengxiang Pan February 13th. Reasons for HTTP Caching. Latency reduction Bandwidth conservation Disconnected operation Not fully support. Typical Tradeoffs. CPU cycles (operational costs) RAM Disk storage
E N D
Squeezing More Bits Out of HTTP Caches Jeffrey C. Mogul Presented by: Zhengxiang Pan February 13th
Reasons for HTTP Caching • Latency reduction • Bandwidth conservation • Disconnected operation • Not fully support
Typical Tradeoffs • CPU cycles (operational costs) • RAM • Disk storage • Implementation complexity & costs Principle: avoid round-trips as possible
Evaluation Methods • Actual trace • Capture stream • Proxy log • Simulation model • Workload generator • Require simplifying assumptions • Both not ideal • Usage pattern variation • Lack of repeatability
Limits of Simple Approaches • Uncachable Resources • Coherency misses • Unique URL for each reference • Compulsory miss • High rate of change • Resource size distribution • possible: small set of small resources accounts for most of the cache hits • HR (hit ratio): 30-49% • BHR/ WHR (byte/ weighted hit ratio): 14-36%
Practical limits • Cache check latency • miss_cost = retrieval_latency + cache_check_latency • Capacity misses • Who faster? Hardware increase or web reference rate grow • Cooperative caching • ->single infinite cache -> not good HRs
Prefetching and Replacement Hints • Prefetching • Make accurate predication • Decide whether sufficient bandwidth exists • Reduce user-perceived latency & increase bandwidth consumption • Replacement Hints • Server supply hints about the replacement strategy
Partial Transfers • HTTP/1.1 range retrieval request • Delta Encoding • Just sending changes to cached • Require server to keep previous instances • Cache-based compaction • Use older cached instances as dictionary for dictionary-based compression • Complex, infeasible? • HTML Macros • Decompose web pages • Require page designer / software support
Alias Discovery and Automatic Duplicate Suppression • A simple and efficient way to detect duplication • MD5? security hole • Alternatives? overhead • HTTP extension to carry meta-data • A new field in HTTP header • A sufficient rate of duplication • 16% in simulation
Simulation results • Add 5.4% to 50.4% HR • Add 6.2% to 32.3% BHR • audio/midi type: 15% duplication ratio • A very small subset of server host produced most duplications • Suggest: send digest or not based on content-type, server host, etc.