NATIONAL TECHNICAL UNIVERSITY OF ATHENS (NTUA)
SCHOOL OF MECHANICAL ENGINEERING
LAB. OF THERMAL TURBOMACHINES
PARALLEL CFD & OPTIMIZATION UNIT

Kyriakos C. Giannakoglou, Professor

Publications
Research Activities
Research Group - BIOs
Development of CFD Methods
Adjoint Methods
Evolutionary Algorithms - EASY
Programming On GPUs
"VELOS" PC Cluster
Courses Taught


The group's research activities comprise the development and use of CFD-based analysis and optimization methods in a variety of application domains. The majority of the routinely performed computations are large scale ones and, thus, computationally demanding. These are all performed on the high-performance platform "VELOS" with 4 clusters and 32 TeraFlop of total computing power. Each cluster consists of a number of interconnected CPUs (nodes) employing the TCP/IP protocol. In particular:
  1. The first cluster consists of 22 nodes (32bit, 35 cores in total). Its total RAM is 20GB, varying from 512MB to 2GB on each node. The operating system is Linux (Fedora).
  2. The second cluster consists of 30 nodes (64bit, 80 cores in total). Its total RAM is 100GB, varying from 1 to 4GB on each node. The operating system is Linux (Fedora and CentOS). Some nodes are equipped with NVIDIA GPU cards (GTX 280, 285 & 580).
  3. The third cluster comprises 42 blade servers (64bit) with 2 Quad core Xeon GPUs each, summing up to 336 cores. The nodes have 8GB or 16 GB RAM, with a total of 544 GB for this cluster. The operating system is Linux (CentOS).
  4. The fourth cluster comprises 4 blade servers (64bit) with 2 Quad core Xeon GPUs and 16 GB RAM each. Each server also has 3 NVIDIA Tesla M2050, 3GB memory each (12 Tesla M2050, 36 GB GPU memory in total). The operating system is Linux (CentOS).
The necessary services required for communication and data storage are handled by four dedicated servers, with the Network Information Service (NIS, holding the necessary user information) and Network File System Service (NFS, for file sharing).

Time consuming numerical simulations employ the MPI and PVM protocols, according to the MIMD (Multiple Instruction Multiple Data) model. Parallel processing with the SIMT (Single Instruction Multiple Thread) model is also performed using Graphics Processing Units (GPUs, NVIDIA's GTX 280, 285, 580 Tesla M2050).

The four clusters of "VELOS" have also been used to run Grid-enabled CFD and optimization software. Regarding Grid Computing, the Grid interconnection is based on 3 layers of middleware, namely GridWay, Globus Toolkit and Condor. GridWay is used as the resource management software on the Grid. It unifies all Grid resources under a common queue and performs matchmaking according to requirements and rank attributes for each job. Once the Grid resource is selected, GridWay posts the job to the Globus Toolkit. This is installed on the front-end of each cluster and provides the framework for the interconnection of remote resources, user authentication, information services and gateways to which jobs are submitted. Globus posts the job to Condor, the local resource managment software of each cluster, which performs matchmaking and job monitoring on this cluster.

Kyriakos C. Giannakoglou,
Professor NTUA, Lab. of Thermal Turbomachines,
School of Mechanical Engineering, National Technical University of Athens,
e-mail: kgianna@central.ntua.gr
Tel. (+30)210-772-1636, Fax (+30)210-772-3789,
Mailing Address: 9, Irron Polytechniou, Polytechnioupolis Zografou, Athens 157 80, GREECE
or: P.O. Box 64069, Athens 15710, Greece

MAIN PAGE