Download presentation
Presentation is loading. Please wait.
Published byTheodore York Modified over 8 years ago
1
System Software Lab. A Scalable Web Cache Consistency Architecture 2000. 10. 17 Kim Sangyup SSLAB. EE. KAIST SIGCOMM ’ 99 Haobo Yu, Lee Breslau
2
System Software Lab. 2 Contents Introduction Previous approaches Time-To-Live Invalidation Lease New Approach Analytic Performance Evaluation Conclusion & Critique
3
System Software Lab. 3 Introduction Web caching Lowers the load on servers Reduces the overall network bandwidth required Lowers the latency of responses But, disadvantage in cache consistency Cache consistency Stale page in the cache is problem Perishable Mark uncachable, expecting “ reload ” button
4
System Software Lab. 4 Previous Approaches Time-To-Live Server allocate TTL on each page To request on expired page send IMS message to server Set TTL to small : provide tight consistency But decrease benefits of web caching TTL = 0 : poll away Adaptive TTL Not give an upper bound on the staleness of a page
5
System Software Lab. 5 Previous Approaches(cont ’ d) Invalidation When page is modified, server send invalidation signal Server keeps state on every client of each page Scaling problems when there are many readers Multicast to transmit the invalidation Assigning a multicast group to each page Other unscalable overhead on the routing infrastructure
6
System Software Lab. 6 Previous Approaches(cont ’ d) Lease Combines TTL and invalidation Whenever stores a page, requests a lease from server Page change -> server notifies all caches(has valid page) Request for expired lease of page -> send IMS message Short lease -> same to TTL Two volume lease algorithm Reduce validation traffic of short lease Page lease : long Volume lease : short
7
System Software Lab. 7 New Approach Multicast-based invalidation on application level Using a cache hierarchy to avoid the scalability problem
8
System Software Lab. 8 New Approach(cont ’ d) Hierarchy Multicast group are associated with caches, not pages Each parent cache owns a unique multicast group Allocating group address Unique sender in the group Heartbeats Keep alive hierarchy Group owner sends out a periodic heartbeat(=t) Same to volume lease of length T(T/t = 5) (current time – highest timestamp) > T : lease expiring
9
System Software Lab. 9 New Approach(cont ’ d) Attaching servers JOIN/LEAVE message Update the server routing table in each cache Top level cache knows about all servers attached in hierarchy Handling request
10
System Software Lab. 10 New Approach(cont ’ d) Invalidations On top of heartbeats, piggyback explicit invalidations Only invalidate pages that have been requested(read page) Each heartbeat contains a list of all read pages that have been rendered invalid the last time period T In child cache Have valid read page->mark invalid and propagate the invalidation Have no valid read page->ignores the invalidation
11
System Software Lab. 11 New Approach(cont ’ d) Adding Push to the Architecture Reduce first access latency Selective push Each cache and server make their own independent decision about whether or not to push a page Receives an invalidate of page P : Request for page P : Push page P In case
12
System Software Lab. 12 Analytic Performance Evaluation Reading & writing : Poisson processed of rate r, w Four event : RR, RW, WR, WW Frequency of event
13
System Software Lab. 13 Conclusion & Critique Conclusion Multicast invalidation with volume leases A Caching hierarchy to make the design more scalable Critique Stable and well-managed cache hierarchy Higher level cache have more overload Make response time longer
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.