Search

A hybrid cluster implies the use of computational devices with radically different architectures. Usually, these are conventional CPU architecture (e.g. x86_64) and GPU architecture (e. g. NVIDIA CUDA). Creating and exploiting such a cluster requires some experience: in order to harness all computational power of the described system and get substantial speedup for computational tasks many factors should be taken into account. These factors consist of hardware characteristics (e.g. network infrastructure, a type of data storage, GPU architecture) as well as software stack (e.g. MPI implementation, GPGPU libraries). So, in order to run scientific applications GPU capabilities, software features, task size and other factors should be considered. This report discusses opportunities and problems of hybrid computations. Some statistics from tests programs and applications runs will be demonstrated. The main focus of interest is open source applications (e. g. OpenFOAM) that support GPGPU (with some parts rewritten to use GPGPU directly or by replacing libraries). There are several approaches to organize heterogeneous computations for different GPU architectures out of which CUDA library and OpenCL framework are compared. CUDA library is becoming quite typical for hybrid systems with NVIDIA cards, but OpenCL offers portability opportunities which can be a determinant factor when choosing framework for development. We also put emphasis on multi-GPU systems that are often used to build hybrid clusters. Calculations were performed on a hybrid cluster of SPbU computing center.

Processing of large amounts of data often consists of several steps, e.g. pre- and post-processing stages, which are executed sequentially with data written to disk after each step, however, when pre-processing stage for each task is different the more efficient way of processing data is to construct a pipeline which streams data from one stage to another. In a more general case some processing stages can be factored into several parallel subordinate stages thus forming a distributed pipeline where each stage can have multiple inputs and multiple outputs. Such processing pattern emerges in a problem of classification of wave energy spectra based on analytic approximations which can extract different wave systems and their parameters (e.g. wave system type, mean wave direction) from spectrum. Distributed pipeline approach achieves good performance compared to conventional "sequential-stage" processing.

Virtual machines are usually associated with an ability to create them on demand by calling web services, then these machines are used to deliver resident services to their clients; however, providing clients with an ability to run an arbitrary programme on the newly created machines is beyond their power. Such kind of usage is useful in a high performance computing environment where most of the resources are consumed by batch programmes and not by daemons or services. In this case a cluster of virtual machines is created on demand to run a distributed or parallel programme and to save its output to a network attached storage. Upon completion this cluster is destroyed and resources are released. With certain modifications this approach can be extended to interactively deliver computational resources to the user thus providing virtual desktop as a service. Experiments show that the process of creating virtual clusters on demand can be made efficient in both cases.

This work describes the idea of the system of informational support for the scientific projects and the development of computational task tracking complex. Due to large requirements for computational experiments the problem of presentation of the information about HPC tasks becomes one of the most important. Nonstandard usage of the service desk system as a basis of the computational task tracking and support system can be the solution of this problem. Particular attention is paid to the analysis and the satisfaction of the conflicting requirements to the task tracking complex from the different user groups. Besides the web service kit used for the integration of the task tracking complex and the datacenter environment is considered. This service kit became the main interconnect between the parts of the scientific project support system and also this kit allows to reconfigure the whole system quickly and safely.