Resource management for scalable disconnected access to Web services

Operating system support for massive replication

“…Gray and Shenoy [10] describe a back of the envelope analysis that compares the dollar value of caching data versus the dollar cost of waiting while the data are refetched at some time in the future [10]; Chandra et. al [6] extend this model to consider prefetching and conclude that using an assumption similar to Gray and Shenoy's estimates of network and disk costs in the year 2000, a system may be economically justified in prefetching an object even if there is only a 1% chance that that object will ever be used. In his Master's thesis [5], Chandra argues that even more aggressive replication could be justified in the future given disk and network cost trends and given the desire to improve end-to-end availability not just performance.…”

“…ii) deployment of large number of services together with aggressive replication by each of them effectively implies near infinite demand for resources. For example, CDNs have to deal with thousands of demanding services and mobile clients that are capable of disconnected services [6], have to deal with tens or hundreds of services all contending for the system resources.…”

“…These challenges will affect the design of operating systems for hosting systems and operating systems for light-weight clients. The Active Names system has been adapted to provide fair self-tuning division of resources across downloaded extension code modules [6]. These techniques could be adapted to partition resources for downloaded code at clients, proxies, and hosting centers.…”

“…Unfortunately, making such services highly available is difficult. For example, even though end servers or service hosting sites advertise an availability of "four nines" (99.99%) or "five nines" (99.999%), the end-to-end service availability (as perceived by clients) is typically limited to two nines because of poor wide area network availability [6]. Moreover, although network bandwidths are improving quickly, network latencies are much more difficult to improve in wide area networks, which limits performance for accessanywhere services if those services are delivered from a single location.…”

The potential costs and benefits of long-term prefetching for content distribution

“…If bandwidth is cheap and human waiting time expensive, prefetching can be justified even if it significantly increases bandwidth demands and only modestly improves response times. For example, Duchamp argues for a prefetch threshold of 0.25 in his hyperlink prefetching system [18], and Chandra et al [12] argue that thresholds as low as 0.01 may be justified given current WAN network transfer costs and human waiting time values.…”