Building a Scientific Cloud Computer with OpenStack

In recent years cloud computing has emerged as an alternative to traditional methods of high-performance computing and as a means to effectively manage physical resources. In 2010 The University of Oregon (UO) was awarded a $2 million Major Research Instrumentation grant to build the Applied Computational Instrument for Scientific Synthesis (ACISS) machine. The goal of ACISS is to provide a high-performance cloud computing and storage infrastructure to support scientific research at the UO College of Arts and Sciences. The multi-disciplinary proposal includes computational projects in Computer Science, Physics, Chemistry, Psychology, and Biology.

ACISS consists of a heterogenous 200 node (2600 core) HP system, with 128×12 core basic nodes, 52×12 core GPU nodes with 156 NVidia 2070 GPUs, 16×32 fat core nodes, and a 400 TB parallel network attached storage (NAS). Networking is handled by a tiered 1 GigE management fabric and a 1 to 1 non-blocking 10 GigE storage and communication fabric. Red Hat Enterprise Linux (RHEL) is the base operating system with OpenStack providing the cloud compute and storage infrastructure.

This talk will cover the building of the ACISS system, with an overview of the data center, computational and storage hardware, and software. Topics will include:

• the methodology that went behind choosing OpenStack as our cloud computing framework, and its advantages and disadvantages as compared to other frameworks such as OpenNebula, Eucalyptus, and Nimbus,
• the challenges faced with integrating OpenStack with RHEL, KVM, high performance NAS, and GPUs,
• integrating OpenStack with the existing read-only campus authentication and directory services,
• the performance of virtualized HPC systems compared to traditional clusters,
• and a brief discussion of the science being conducted on ACISS.