Kotlin Multiplatform & Compose Multiplatform - Starter kit for pragmatics
Exploiting Availability Prediction in Distributed Systems
1. Exploiting Availability
Prediction in Distributed
Systems
James W. Mickens and Brian D. Noble
Review by Mario Almeida (EMDC)
Problem statement
What is the problem addressed?
Distributed Systems have scalability and cost advantages, but also introduce issues such
as availability of nodes. This paper presents techniques for predicting availability such as
availability-guided replica placement, improvement routing in delay-tolerant networks and
combining availability prediction with virus modeling. It claims that this way one can reduce
overheads by predicting availability and planning for changing availability instead of reacting to
it.
Why is it important?
Because availability can be crucial for keeping a service functionality as well as replication
of contents. Availability of nodes can have a direct impact on the overall performance of any
distributed system, such as churn that can lead to significant overheads.
Proposal
What is the proposed solution?
The paper proposes the definition of new techniques for predicting availability. They used
multiple predictors such as the RightNow Predictor, SatCount Predictor, state-based predictor,
TwiddledHistory Predictor, Linear Predictor and a Hybrid predictor that changes predictor
depending on lookahead periods.
2. Hypotheses
What were the expected effects of the proposed solution?
The paper describes three applications using availability prediction. They expected the first
application, a modified DHT with availability-aware replica placement, to transmit fewer objects
for regeneration and to have a better data availability. The second application to have a better
routing performance and the last to combine virus modeling with availability predicting for
achieving better forecasts.
Experiments
What were the experiments?
The hybrid predictor was tested using PlanetLab and Microsoft traces. They also used overnet
traces to model uptime patterns.
In order to experiment on decentralized availability data, the three case studies mentioned
before were used :
● simulation of Chord DHT in order to see how availability-guided data placement reduces
copy overhead by using replica sites;
● message delay in delay-tolerant networks;
● combine virus modeling with availability predicting for achieving better forecasts.
Results
What were the results?
The availability predictors introduced in this paper have shown that the overhead can be
reduced by planning of changing availability. In the first case test, by using replica storage,
bandwidth consumption was reduced and data availability was increased. For the second case,
latencies in the delay-tolerant network were decreased by using availability predictors. And
finally, forecast of global infection is improved by combining availability predictions with virus.