A. Fox and E. A. Brewer. Harvest, yield and scalable tolerant systems. In Proceedings of the 1999.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....or reference version. Data fidelity, while valuable, is not sufficient to describe the entire range of fidelity adaptation. Consider a search over an image database. We could run an approximate version of the query, which runs faster but might be more inaccurate: e.g. we can reduce the harvest [36] the amount of data examined as in online aggregation [46] or congressional sampling [4] Another example of approximate querying is approximate medians [60] where accuracy is traded for main memory usage. In all these cases, we degrade fidelity by providing a less trustworthy result: ....

Armando Fox and Eric A. Brewer. Harvest, yield and scalable tolerant systems. In Proceedings of the 7th Workshop on Hot Topics in Operating Systems (HotOS-VII), pages 174--178, Rio Rico, AZ, March 1999. IEEE Computer Society.

.... seem to be the first directed to studying these servers, although recent studies have looked at response time and availability of a single node Apache Web server [18] Though other previous work proposes availability improving strategies for applications spanning large configurations [9], we seem to be the first group to quantify the performability and the design and environmental tradeoffs of cluster based servers. Finally, we recently used the methodology introduced here to quantify the effects of two different communication architectures on the performability of PRESS [25] ....

A. Fox and E. Brewer. Harvest, Yield and Scalable Tolerant Systems. In Proceedings of Hot Topics in Operating Systems (HotOS VII), Mar. 1999.

....23, 18, 24] are necessary background for a model based quantification effort such as ours. Our methodology and infrastructure seem to be the first directed to quantifying the availability impact of a range of techniques as applied to cluster based services. One of the first works on the subject [13] argued that there are irresolvable tradeoffs between availabil ity, consistency, and performance in these services, but did not quantify the impact of these tradeoffs. A more recent work developed models which can be used to quantify the impact of faulty components on quality of service metrics ....

A. Fox and E. Brewer. Harvest, Yield and Scalable Tolerant Systems. In Proceedings of Hot Topics in Operating Systems (HotOS VII), Rio Rico, AZ, Mar. 1999.

....provide much needed data on actual fault profiles. Works in the past have proposed robustness [25] and reliability benchmarks [29] that quantify the degradation of system performance under failures. Previous work has noted that different cluster organizations have different availability impacts [11]. Our work focuses on a methodology that encompasses both. Mendosus is related to a number of efforts that have explored hardware and software based fault injection techniques [10, 13, 17, 24] Mendosus uses an approach similar to Orchestra [10] which uses a software based, script driven ....

....a fault model comprised of higher level faults such as a network partition and mapping it into the low level faults that Mendosus can inject is one direction. In addition, we will extend the metrics used to include latency and utilization, as well as more novel metrics such as harvest and yield [11]. A longer term modeling goal is to examine the impact of non exponential fault distributions, which would allow us to model components that degrade over time. Longer term, we will explore the tradeoffs between performance, availability, and redundancy. Extra capacity can be used for either ....

A. Fox and E. Brewer. Harvest, Yield and Scalable Tolerant Systems. In Proceedings of Hot Topics in Operating Systems (HotOS VII), Rio Rico, AZ, Mar. 1999.

....providing much needed data on actual fault profiles. Works in the past have proposed robustness [25] and reliability benchmarks [29] that quantify the degradation of system performance under failures. Previous work has noted that different cluster organizations have different availability impacts [11]. Our work focus on a methodology that encompasses both. Mendosus is related to a number of efforts that have explored hardware and software based fault injection techniques [10, 13, 17, 24] Mendosus uses an approach similar to Orchestra [10] which uses a software based, script driven fault ....

....a fault model comprised of higher level faults such as a network partition and mapping it into the low level faults that Mendosus can inject is one direction. In addition, we will extend the metrics used to include latency and utilization, as well as more novel metrics such as harvest and yield [11]. A longer term modeling goal is to examine the impact of non exponential fault distributions, which would allow us to model components that degrade over time. Longer term, we will explore the tradeoffs between performance, availability, and redundancy. Extra capacity can be used for either ....

A. Fox and E. Brewer. Harvest, Yield and Scalable Tolerant Systems. In Proceedings of Hot Topics in Operating Systems (HotOS VII), Rio Rico, AZ, Mar. 1999.

....under realistic load in the proving ground pool is necessary regardless of the cause for taking a machine out of active service. 4.2 Evaluation Metric Before evaluating our validation based policy, we first consider the question of which metric to use. We originally used yield, first proposed in [9], which is the percentage of successful requests. Yield, however, turned out to be too coarse of a metric. In particular, yield is very sensitive to the time interval over which it is measured. For example, a service which loses 100 requests evenly spread over 100 hours is much different than one ....

....computed as the number of reboots that occur in the proving ground pool divided by the total number of reboots. As we scale load, we can generate an avoidance curve. Armed with an avoidance curve and enough data about actual numbers of reboots, a designer can compute other metrics such as harvest [9], yield and availability given a data partitioning scheme and the parametric assumptions in Table 3. 4.3 Simulation Model We use a simple abstract model of cluster based Internet services, much like the one in [2] The model is comprised of a frontend load director and a cluster of back end ....

A. Fox and E. Brewer. Harvest, Yield and Scalable Tolerant Systems. In Proceedings of Hot Topics in Operating Systems (HotOS VII), Rio Rico, AZ, Mar. 1999.

....built on top of them are to be highly available. In addition, because Internet services can have widely different persistence and or consistency requirements, our data structures will need to leverage the consistency and persistence properties of specific applications to ensure good performance [5, 13, 14, 21]. Preliminary Work: On Line Sorted Lists We are in the process of implementing our first data structure, an on line sorted list that allows fast look ups, insertions and deletions. Each item in the list is a key value pair. Items in the list are maintained in sorted order according to their ....

A. Fox and E. Brewer. Harvest, Yield and Scalable Tolerant Systems. In Proceedings of Hot Topics in Operating Systems (HotOS VII), Rio Rico, AZ, Mar. 1999.

....built on top of them are to be highly available. Finally, because Internet services can have widely different persistence and or consistency requirements, our data structures will need to leverage the consistency and persistence properties of specific applications to ensure good performance [13, 5, 12, 21]. Preliminary Work: On Line Sorted Lists We are in the process of implementing our first data structure, an on line sorted list that allows fast look ups, insertions and deletions. Each item in the list is a key value pair. Items in the list are maintained in sorted order according to their ....

A. Fox and E. Brewer. Harvest, Yield and Scalable Tolerant Systems. In Proceedings of Hot Topics in Operating Systems (HotOS VII), Rio Rico, AZ, Mar. 1999.

....with our metrics. Availability is not explicitly addressed in [21, 22] In two recent papers[8, 29] metrics similar to Unseen Writes and Staleness are used to measure database freshness. However, no design is proposed to provide guaranteed freshness levels by bounding the metrics. Fox and Brewer[15] argue that strong consistency (onecopy serializability[4] and one copy availability[32] cannot be achieved simultaneously in the presence of network partitions. In the context of the Inktomi search engine, they show how to trade harvest for yield. Harvest measures the fraction of the data ....

Armando Fox and Eric Brewer. Harvest, Yield and Scalable Tolerant Systems. Proceedings of HotOS-VII, March 1999.

....for example, the Web server degrades the compression quality of JPEG images when bandwidth utilization exceeds a target. In some cases it is possible for a service to make performance tradeoffs in terms of the freshness, consistency, or completeness of data delivered to clients. Brewer and Fox [44] describe this tradeoff in terms of the harvest and yield of a data operation; harvest refers to the amount of data represented in a response, while yield (closely related to availability) refers to the probability of completing a request. It is often possible to achieve better performance from an ....

A. Fox and E. A. Brewer. Harvest, yield and scalable tolerant systems. In Proceedings of the 1999.

No context found.

A. Fox and E. A. Brewer. Harvest, yield and scalable tolerant systems. In Proceedings of the 1999.

No context found.

A. Fox, , and E. Brewer. Harvest, yield and scalable tolerant systems. In Proceedings of the Seventh Workshop on Hot Topics in Operating Systems, Mar. 1999.

No context found.

Armando Fox and Eric A. Brewer. Harvest, Yield and Scalable Tolerant Systems. In Proceedings of the Seventh Workshop on Hot Topics in Operating Systems, pages 174--178, 1999.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC