A Square Root Topologys To Find Unstructured Peer-To ... - ijcsmr

More documents

Recommendations

Info

International Journal of Computer Science and Management Research Vol 2 Issue 3 March 2013ISSN 2278-733Xevaluated keyword queries against the web pages.Then we generated 20,000 synthetic queriesmatching 631,320 synthetic documents, stored at20,000 peers, such that the statistical properties ofour synthetic content model matched those of thereal trace. The resulting content model allowed usto simulate a network of 20,000 peers. In thissimulation, we submitted random queries chosenfrom the set of 20,000 to produce a total of 100,000query submissions. we describe the details of thismethod of generating synthetic documents andqueries, and provide experimental evidence that thecontent model, though synthetic, results in highlyaccurate simulation results. The synthetic modelretains an accurate distribution of the popularity ofpeer content, which is critical for the construction ofthe square-root topology.3.2 Random walksWe conducted an experiment to examine theperformance of random walk searches in differenttopologies. This queries matched documents storedat different peers, and had a goal G = 10 results. Wecompared three different topologies.• A square-root topology, generated by assigning adegree to each peer based on equation (2), and thencreating links between randomly chosen pairs ofpeers based on the assigned degrees.• A low-skew power-law topology, generated usingthe PLOD algorithm . In this network, α = 0.58.• A high-skew power-law topology, generated usingthe PLOD algorithm, with α = 0.74.Random walks in the square-root topology require8,940 messages per search, 26 percent less. thanrandom walks in the low-skew power-law topologyand 45 percent less than random walks in the highskewpower-law topology. In the power-lawtopologies, searches tend toward high degree peers,even if the walk is truly random and not explicitlydirected to high degree peers. These high degreepeers also have the most popular content, Result isthat searches have a low probability of going to thepeer with matching content, and the number of hopsand thus messages increases. If the power-lawdistribution is more skewed, then the probabilitythat searches will congregate at the wrong peers ishigher and the total number of messages arenecessary to get to the right peers increases. Eventhough random walks perform best in the squareroottopology, a large number of messages need tobe sent. the result is a significant improvement overtraditional Gnutella style search, flooding in a highskewpower-law network, with a TTL of five inorder to find at least ten results on average, requires17,700 messages per search. The above results arefor simple, unoptimized random walks. Addingoptimizations such as proactive replication orneighbor indexing reduces the cost of a randomwalk search, and results for these techniques showthat the square-root topology is still best. Anotherissue with random walks is that the search latency ishigh, as queries may have to walk many hopsbefore finding content. To deal with this, Lv et alpropose creating multiple, parallel random walksfor each search. Since the network processes thesewalks in parallel, the result is reduced searchlatency. We ran experiments where we created 2,10,15, 20, 30, and 100 parallel random walks for eachsearch, and measured search latency.WalksSquarerootPower-lawPower lawlow-skewhigh-skew1 8930 12090 163502 4500 6210 89705 1800 2490 374010 904 1250 188020 454 630 947100 96 130 194Table2. Parallel random walks: search latency(ticks).These results are shown in Table 2.Thesquare-root topology provided the lowest searchlatency, regardless of the number of parallel walksthat were generated. The improvement for thesquare-root topology was consistently 27 percentcompared to the low-skew power-law topology, and50 percent compared to the high-skew power-lawtopology. Even when searches are walking inparallel, the square root topology helps those searchwalks quickly arrive at the peers with the rightcontent.3.3 Proactive ReplicationThe square-root topology is complementary to thesquare-root replication. It is feasible to proactivelyreplicate content, the square-root replicationspecifies that the number of copies made of contentshould be proportional to the square root of thepopularity of the content. The square-root topologycan be used whether or not proactive replication isused, the combination of the two techniques canprovide significant performance benefits. Weconducted an experiment where Replicated contentaccording to the square-root replication. Then weconnected peers in the square-root, high-skewpower-law, and low-skew power-law topologies,and states the performance of random walksearches. Again, G= 10. As expected, proactiveD. Raman et.al. 1734www.ijcsmr.org
International Journal of Computer Science and Management Research Vol 2 Issue 3 March 2013ISSN 2278-733Xreplication provided better performance than noreplication. Proactive replication performs best withthe square-root topology, requiring only 2,830messages per search, 42 percent less than in thelow-skew power-law network and 56 percent lessthan in the high-skew power-law network.replication makes more copies of the documentsthat a search will match, while the square-roottopology makes it easier for the search to get to thepeers where the documents are stored. Thecombination of the two techniques provides moreefficiency than either technique alone. For example,the square root topology with proactive replicationrequired 68 percent fewer messages than the squareroot topology without replication.3.4 Other search walk techniquesWe examined the performance of other walk-basedtechniques on different topologies. We comparedthree other techniques based on random walks:• Biased high degree messages are preferentiallyforwarded to neighbors that have the highestdegree.• Most results messages are forwardedpreferentially to neighbors that have returned themost results for the past 10 queries.• Fewest result hops messages are forwardedpreferentially to neighbors that returned results forthe past 10 queries who have travel the fewestaverage hops .In each case, ties are broken randomly. For thebiased high degree technique, we examined bothneighbour indexing and no neighbour indexing.Although describes several ways to route searchesin addition to most results and fewest result hops,these two techniques represent the best that theresult hops requires the least bandwidth, while mostresults has the best chance of finding the requestednumber of matching documents. The square-roottopology is best. The most improvement is seenwith the biased high degree technique, where theimprovement on going from the high-skew powerlawtopology to the square-root topology is 52percent. Large improvements are achieved with thefewest result hops technique and most results. Thesmallest improvement observed was for the biasedhigh degree technique with neighbor indexing.Square-root topology offers a 16 percent decrease inmessages compared to the lows-kew power-lawtopology. The square-root topology provides thebest performance, even with the extremely efficientbiased high degree / neighbor indexingcombination. The square-root topology can be usedeven when neighbor indexing is not feasible. Thecombination of square-root topology, square-rootreplication and biased high degree walking withneighbor indexing provides even betterperformance. The Results indicate that thisapproach is extremely efficient, requiring only 248messages per search on average. The square roottopology is better than the power law topologywhen square-root replication and neighbor indexingare used. Using three techniques together results ina searching mechanism that contacts less than 2percent of the systems peers on average while stillfinding sufficient results. The results so far assumestate-keeping, where peers keep state about wherethe search has been. Peers can avoid forwardingsearches to neighbours that the search has alreadyvisited. The results demonstrate that the square-roottopology is better than power-law top logies,whether or not state keeping is used.3.5 Other TopologiesWe also tested the square-root topology incomparison to several other network structures. wecompared against two simple structures.• Constant-degree topology: every peer has thesame number of neighbors. In our simulations, eachpeer had five neighbors.• Proportional topology: every peer had a degreeproportional to their popularity gk.Our results show that the square-root topology isbest, requiring 10 percent fewer messages than theconstant degree network, and 7 percent fewermessages than the proportional topology. Althoughthe improvement is smaller than when comparingthe square-root topology to power-law topologies,these results again demonstrate that the square-roottopology is best. The cost of maintaining thesquare-root topology is low, as we discuss inSection 4, requiring easily obtainable localinformation. It clearly makes sense to use thesquare-root topology instead of constant degree orproportional topologies. A widely used topology inmany systems is the super-peer topology. In thistopology, a fraction of the peers serve as superpeers,aggregating content information from severalleaf pears. Then, searches only need to be sent tosuper-peers. They are connected using a normalunstructured topology. We ran simulations using astandard super peer topology, in which searches areflooded to super peers. We compared this standardtopology to a super peer topology that used thesquare-root topology and random walks betweensuper peers. The results indicate a significantimprovement using our techniques the square rootsuper peer network required 54 percent fewermessages than a standard super-peer network.4 ConclusionsD. Raman et.al. 1735www.ijcsmr.org
Page 2 and 3: International Journal of Computer S
Page 6: International Journal of Computer S

A Square Root Topologys To Find Unstructured Peer-To ... - ijcsmr

You also want an ePaper? Increase the reach of your titles

Delete template?

Save as template?