ICTTA'04Arwa zabian1 On The Latency of BFS Interval Cooperation Web Caching Arwa Zabian Maurizio Bonuccelli Department of Computer Science University of Pisa / Italy

ICTTA'04Arwa zabian2 Talk overview Give a brief introduction of the problem The presentation of our algorithm Our results Conclusion and future works

ICTTA'04Arwa zabian3 Client Web caching

ICTTA'04Arwa zabian4 Problems of Web Caching Problems of Web Caching Limited cache size Cosistency problem Caching dynamic content Solutions Prefetching Cooperative Web Caching

ICTTA'04Arwa zabian5 Cooperative web caching

ICTTA'04Arwa zabian6 Problems in cooperative web caching Where the request must be forwarded? How it can be retrieved?

ICTTA'04Arwa zabian7 Goal Goal The goal is the reduction of the latency perceived by the web client. Delay analysis in a cooperative web caching: -N-Network factors -S-System factors Delay Searching Finding the proxy Findin R Routing decision Fault Tolerance System factors

ICTTA'04Arwa zabian8 Delay motivations: Searching Defining the topology Finding the proxy that have the request, it depends on the topology of the system. The proxy cache in our system are organized in a tree. The tree construction is done using NDA (Neighbour Discovering Algorithm). NDA is based on BFS, it visits the tree by levels.

ICTTA'04Arwa zabian9 Defining the topology NDA (Neighbours Discovering Algorithm)

ICTTA'04Arwa zabian10 Delay motivation: Seaching Finding the request When NDA runs in a given graph it constructs a tree called Child-Tree, in which each node reports its parent label. The maximum length of label assigned to n nodes is 2logn+1 bits. Based on the information collected each node constructs a routing table that is a data structure in which is inserted the node labels and their Interval of documents. The documents are organized in intervals of their URL based on the interval routing scheme proposed by Santoro and Khatib[82]. N. Santoro, R. Khatib: Routing Without Routing Table. Tech. Rep, SCS-TR-6 School of Computer Science, Carleton University. Ottawa 1982

ICTTA'04Arwa zabian11 Searching : Complexity analysis NDA is asymptotically optimum as search graph algorithm. Time complexity is O(|E|+|V|). Message complexity in construction Child-Tree is O(|E|+|V|).

ICTTA'04Arwa zabian12 NDA contribues in the routing process in two factors they are: Routing decision. Address solution at each hop. Fault tolerance: it maintains a backup links used in the case of nodes and links faulty. Delay motivation Routing and fault tolerance

ICTTA'04Arwa zabian13 Routing : Complexity analysis Routing : Complexity analysis The size of data structure constructed by the information collected during the NDA discovering process is logarithmic with the number of nodes of the graph. The number of nodes travelled by each message, in Child- Tree is O(log n).

ICTTA'04Arwa zabian14 Simulation Results Simulation Results All our simulation were run on top of NS2. The goal was the reduction of the searching and routing time. Our simulation results confirm that the searching time is fixed in all the conditions studied and varied only in the case of miss. The routing time is strictely related to the size of the document and to the number of hops between proxy and server

ICTTA'04Arwa zabian15 Simulation Results: factors influenced in the latency Type of the document: static or dynamic. Page popularity. Number of request. Number of hops. Page size

ICTTA'04Arwa zabian16 Type of the document : static or dynamic Type of the document : static or dynamic Static Dynamic

ICTTA'04Arwa zabian17 Page Popularity Page Popularity Static and popular page Static and less popular page

ICTTA'04Arwa zabian18 Number of request Number of request The latency is decreased when the number of requests is increased between two modification.

ICTTA'04Arwa zabian19 Simulation Results: Number of hops Simulation Results: Number of hops The retrieving process R R R

ICTTA'04Arwa zabian20 Simulation Results: Number of hops Simulation Results: Number of hops The latency is increased linearly with the number of hops

ICTTA'04Arwa zabian21 Simulation Results : Page size Simulation Results : Page size The simulation results show that the most important factor that influences in the latency variation is the page size. The latency is increased linearly with the page size. Hierarchical Randomized

ICTTA'04Arwa zabian22 Simulation Results: Page size Simulation Results: Page size The hit ratio is influenced by the page size. 100% 18% 82% 82%

ICTTA'04Arwa zabian23 Simulation Results: Page size The transmission time is strictly related to the page size that affect the latency. The transmission time is increased linearly with the page size Page size (bytes) T.T (ms) Latency (ms) Randomised Latency(ms) Hierarchical

ICTTA'04Arwa zabian24 Simulation Results : Transmission Time T.T hierarchical randomised

ICTTA'04Arwa zabian25 Results: Results: Scalability: the latency is independent on the network size Number of nodes Average of Latency (ms) Randomised Average of Latency (ms) Hierarchical

ICTTA'04Arwa zabian26 Results : Scalability

ICTTA'04Arwa zabian27 Results: Results: Reliability : the system ensures that the client receives the request in the presence of some links faults with some additional delay penalty. With faultWithout fault Average of latency (ms) for 2 hops Average of latency (ms) for 6 hops

ICTTA'04Arwa zabian28 Related works Related works Riptide is a peer-to-peer web caching system built on top of OceanStore that uses a routing mechanism called Tapestry. The comparison of our system with that for Riptide show that : our system perform well for small documents. Both systems give a similar results for medium document. When Riptide performs better than our system for large documents

ICTTA'04Arwa zabian29 Related works Comparison results in term of latency Latency(ms) Riptide Latenc(ms) Hierarchical NDA Latency (ms) Randomised NDA Page size(bytes)

ICTTA'04Arwa zabian30 Related work Comparison results in term of hit 100% Riptide NDA 100% 82%

ICTTA'04Arwa zabian31 Conclusion Conclusion We proposed an algorithm based on BFS that reduces the searching time by reducing the size of the data structure used rapresenting the documents in intervals based on their URL’s. Our algorithm was proposed to reduce the routing time by reducing the address resolution in each node. We obtained a reliable system that allow the client to receive the request with the presence of some faulty link in the system. It is scalable in which the latency perceived by the client is independent on the system size.

ICTTA'04Arwa zabian32 Future work Future work Utilizing the document division method in developing a technique for the request prefetching between proxy and proxy. Modelling the system to include the dynamic case ( insertion and deletion of nodes). Performs some additional simulations to include the cases not covered in our previous work.

ICTTA'04Arwa zabian33 Questions?

ICTTA'04Arwa zabian34 References [1] A. Zabian, M. Bonuccelli ``BFS Based Algorithm for Routing and Fault Tolerance in a Cluster of Web Caching''. Technical Report, `` ~ bonucce''. [2] B. Zaho, A.Joseph and J. Kubiatowicz, ''Tapestry: Infrastructure For Fault-tolerance Wide Area Location and Routing''. Technical Report UCB//CSD , U.C. Berckley [3] C. Wells. ``The Oceanstore Archive: Goal, Structure and self Repair'', Master thesis, University of California, Berkeley.May2001 [4] Duane Wessels.``Internet Cache Protocol(ICP)'', version 2. RFC [5] E. Zegura, K. Calvert and S. Bhattacharjee. ''How to model an Internetwork''. In Proc. of INFOCOM, [6] L.Fan, P. Cao, J. Almedia, A.Z. Border,``Summary-Cache a scalable Wide Area Web Cache sharing Protocol'', IEEE/ACM Transactions on Networking. Vol 8. No 3,pp: June [7] Martin Hamilton, Alex Rousskov, Duane Wessels, ``Cache Digest'', National Laboratory for Applied Network Research, Aprile1998. “ [8] Patrick R.Eaton ``Caching the web with OceanStore'‘. Technique report UCB/CSD/ U.C.Berkely. November 2002.