System X: building the Virginia Tech supercomputer

Author

Varadarajan, Srinidhi

Author_Institution

Terascale Comput. Facility, Virginia Polytech. Inst. & State Univ., Blacksburg, VA

fYear

2004

fDate

11-13 Oct. 2004

Firstpage

2

Abstract

System X was conceived in March 2003, designed in July 2003, and by October it had achieved a sustained performance of 10.28 Teraflops, making it the third fastest supercomputer in the world today. System X has several novel features. First, it is based on an Apple G5 platform with the new IBM PowerPC 970 64-bit CPUs. Secondly, it uses a high performance switched communications fabric called Infiniband. Finally, system X is cooled by a hybrid liquid-air cooling system. In this paper, the author presents the motivation for System X, its architecture, and the challenges faced in building, deploying, and maintaining a large-scale supercomputer. The paper is focused on transparent fault tolerance for massively parallel supercomputers, scalable network emulation, compiler directed strategies for flexible data sharing models, and routing algorithms for backbone IP networks

Keywords

IP networks; fault tolerance; learning (artificial intelligence); mainframes; parallel machines; routing protocols; telecommunication traffic; AI technique; Infiniband; System X; Virginia tech supercomputer; backbone IP network; checkpointing algorithm; compiler directed strategy; hybrid liquid-air cooling system; large-scale supercomputer; migration algorithm; multipath routing protocol; network traffic visualization; parallel supercomputer; reinforced learning; routing algorithm; scalable network emulation; switched communications fabric; transparent fault tolerance; Buildings; Communication switching; Cooling; Emulation; Fabrics; Fault tolerance; Large-scale systems; Routing; Spine; Supercomputers;

fLanguage

English

Publisher

ieee

Conference_Titel

Computer Communications and Networks, 2004. ICCCN 2004. Proceedings. 13th International Conference on

Conference_Location

Chicago, IL

ISSN

1095-2055

Print_ISBN

0-7803-8814-3

Type

conf

DOI

10.1109/ICCCN.2004.1401571

Filename

1401571