Wednesday, 21 March 2012

Belgium VPN between the NaN

connectivity line represents a system where each node storesthe 233 most popular movies (i.e. they only fetch from theirlocal disk or from the ISP and use no connectivity to theirneighbors). The cooperative file caching line represents a system where files (movies) are sorted in terms of popularity andeach file, in order, is assigned to a node in the neighborhooduntil all space is exhausted. This represents a cooperativecaching of web content that is connectivity-unaware andagnostic to streamed content. This places as much content inthe neighborhood without considering the limitations of localconnectivity. The w/ NaN line represents our proposed contentplacement that considers both popularity and connectivity.There are several noteworthy observations from this simulation. First, pushing content to the neighborhood providessignificant benefits – all designs do much better than w/oNaN. Second, being popularity aware is critical. Popularityunaware placement (simply placing 11% of the library—1111randomly chosen movies— not shown on graph) results inonly a 25% improvement over w/o NaN. Third, the fact thatNaN significantly outperforms no connectivity shows takingadvantage of connectivity provides significant benefit. Fourth,the difference Belgium VPN between the NaN and cooperative file cachinglines show that the system must monitor neighborhood connectivity and place content carefully considering the streamabilityrequirement. Our approach (the NaN line) reduces the peakbandwidth demand at the second-mile link by 42% (from1980Mbps to 1143Mbps), the 95 percentile bandwidth by 45%(from 1890Mbps to 1037Mbps), and the average bandwidthby 45% (from 1210Mbps to 670Mbps). Note that this verifiesour assertion that the peak and average savings are likely tobe similar in larger neighborhoods.2) Placement Robustness and Fault Tolerance: The transferservice masks failure of local resources by fetching unavailablechunks from the ISP's server. This results in degraded performance. We now evaluate the system when there are contentpopularity mispredictions, unexpected node failures, and linkfailures. Here, we are interested in performance assuming thatthe system does not re-run its placement algorithm to find anew placement. Our results show that the system performanceis not overly sensitive to such failures

No comments:

Post a Comment

Note: only a member of this blog may post a comment.