From SONGS: Simulation Of Next Generation Systems
- Postdoc positions at Inria Bordeaux
- Periodic service allocation for PaaS Clouds
With the increased use of PaaS Cloud Computing platforms to easily
deploy elastic online services, the energy and resource usage of the
very large datacenters which host these platforms is getting
massive. Consolidation and co-allocation are possible solutions to
alleviate this, but this requires efficient allocation algorithms to
cope with the complexity of the problem. In particular, the dynamicity
of the demand from the users of these online services and the quality
of service constraints required by the operators are important factors
which need to be taken into account. Due to the sheer size of the
platforms, the available solutions are mostly simple greedy strategies
which can not provide long-term vision of the allocation.
A very promising direction to improve the efficiency of allocations is
to make use of the observed periodicity of the resource demand, which
exhibit very strong daily and weekly patterns. The objective of this
postdoc is to propose and evaluate efficient allocation algorithms in
this context of periodic demands, with the objective to minimize the
energy and resource usage of the platforms, while also keeping the
amount of dynamic modifications (through migrations / reallocations)
as low as possible.
The solutions proposed will be analyzed analytically and/or with
extensive simulations using the SimGrid framework.
- Modeling and Analysis of dynamic Map Reduce schedulers
The tremendous increase in the size and heterogeneity of supercomputers and cloud platforms makes it very difficult to predict the performance of a scheduling or a resource allocation algorithm. Therefore, dynamic solutions, where scheduling/alllocation decisions are made at runtime have overpassed static strategies. The simplicity and efficiency of dynamic schedulers such as Hadoop are a key of the success of the MapReduce framework.
In this context, it is crucial to understand the influence of the parameters of the dynamic scheduler (in the case of Hadoop for instance the number of replicas, their placement, the dynamic allocation strategy) through the analysis of mathematical models. The first part of the postdoc will therefore be devoted to the modeling of Hadoop algorithm and its influence on the design of the algorithms.
Up to now, the use of MapReduce is restricted to applications that can be written as a sequence of map and reduce operation and therefore who operate at each step on independent tasks and data. In a second step, our goal will be, following [HPDC14] to propose a more general model (in particular suitable to linear algebra operations) and to understand how data placement and task allocation strategies can be adapted to this new context.
Throughout all the postdoc, conclusions will be assessed through an extensive use of simulations using the SimGrid framework
[HPDC14] O. Beaumont and L. Marchal. Analysis of Dynamic Scheduling Strategies for Matrix Multiplication on Heterogeneous Platforms. 23rd ACM Symposium on High-Performance Parallel and Distributed Computing