IT resources to Virtual Infrastructures (VIs) (i.e. groups of VMs, virtual switches, and their network interconnections) is an NP-hard problem. Most allocation algorithms designed to run on CPUs face scalability issues when considering current cloud data centers comprising thousands of servers. This work offers and evaluates a set of allocation algorithms refactored for Graphic Processing Units (GPUs). Experimental results demonstrate their ability to handle three large-scale data center topologies.

Large clusters and supercomputers are rapidly evolving and may be subject to regular hardware updates that increase the chances of becoming heterogeneous. Homogeneous clusters may also have variable performance capabilities due to processor manufacturing, or even partitions equipped with different types of accelerators. Data distribution over heterogeneous nodes is very challenging but essential to exploit all resources efficiently. In this article, we build upon task-based runtimes' flexibility to study the interplay between static communicationaware data distribution strategies and dynamic scheduling of the linear algebra LU factorization over heterogeneous sets of hybrid nodes. We propose two techniques derived from the state-of-the-art 1D×1D data distributions. First, to use fewer computing nodes towards the end to better match performance bounds and save computing power. Second, to carefully move a few blocks between nodes to optimize even further the load balancing among nodes. We also demonstrate how 1D×1D data distributions, tailored for heterogeneous nodes, can scale better with homogeneous clusters than classical block-cyclic distributions. Validation is carried out both in real and in simulated environments under homogeneous and heterogeneous platforms, demonstrating compelling performance improvements.

No abstract

Parallel applications performance strongly depends on the number of resources. Although adding new nodes usually reduces execution time, excessive amounts are often detrimental as they incur substantial communication overhead, which is difficult to anticipate. Characteristics like network contention, data distribution methods, synchronizations, and how communications and computations overlap generally impact the performance. Finding the correct number of resources can thus be particularly tricky for multi-phase applications as each phase may have very different needs, and the popularization of hybrid (CPU+GPU) machines and heterogeneous partitions makes it even more difficult. In this paper, we study and propose, in the context of a task-based GeoStatistic application, strategies for the application to actively learn and adapt to the best set of heterogeneous nodes it has access to. We propose strategies that use the Gaussian Process method with trends, bound mechanisms for reducing the search space, and heterogeneous behavior modeling. We compare these methods with traditional exploration strategies in 16 different machines scenarios. In the end, the proposed strategies are able to gain up to ≈51% compared to the standard case of using all the nodes while having low overhead.