|
|
|
| e-Pub |
Previous | 
Home
Section: New Results
Transparent Resource Management for Clouds
Participants : Julien Gossa, Rajni Aron, Stéphane Genaud, Étienne Michon, Marc-Eduard Frîncu.
Provisioning strategies.
Our main achievement was the design of one comprehensive provisioning meta-strategy. This meta-strategy only use one parameter as a deadline given by the user. Contrary to other deadline-based provisioning strategies, our meta-strategy is able to combine any provisioning strategy in order to optimize the cost while meeting the deadline. This is achieved through simulation of cost and makespan of every available strategy thanks to SCHIaaS 5.4.3 . It allows to apply the most inexpensive strategy as long as possible, before progressively switching to more expensive strategy when the deadline becomes closer.
The next step is to asses this meta-strategy among an important set of applications and platforms, both in real environments and simulation. The data are currently gathered and analyzed, and we should be able to draw conclusions soon.
User workload analysis.
We have conducted one broad study about workflows execution on the cloud, from both the theoretical and experimental point of view. In this study, we tried to discover causalities between the characteristics of workflows and the performances of provisioning strategies. We concluded that, except very peculiar cases, no causality can be identified. That is why we decided to make use of simulation to predict the strategies performances.
This predictive process is now integrated as a module of our cloud broker. It can be invoked by a user to help him decide which strategy should be used before any actual resource leasing.
We are now convinced that workload analysis is not a suitable approach because of its lack of generality.
Experimentations.
Given the very large consumption of CPU hours, the above work was supported mostly by simulation. We have assessed the gap between the performances of real executions on a private cloud and simulation. The latter proved to be very accurate, predicting almost perfectly the cost and makespan of every strategy on a wide range of workloads.
However, we have also shown that the simulation can be very sensitive to user defined input parameters (such as task runtimes) and may be mislead in borderline cases. Identifying the pitfalls and limitations of the simulation is very important and should end up in recommendations for a wise interpretation of simulation results.
We have also extended the range of experimentations to assess our simulator. First, we have extended the set of simulations with new applications, mostly workflows that are both generated and real applications (i.e. Montage). Second, we have conducted intensive experimentations on new platforms (i.e. Bonfire). The experimental data we have recently gathered in both cases is to be analyzed to further validate our approach.