NASA Case Study

ManualsBrandsIntel ManualsOtherIntel 10 Gigabit XF SR Server Adapter

NASA’s Flexible Cloud Fabric:

Moving Cluster Applications to the Cloud

Network architects at the NASA Center for Climate Simulation recently

began investigating the viability of running the organization’s modeling

and simulation applications on cloud infrastructure, as an alternative to

its purpose-built computing cluster named Discover. Hoping to capture

the inherent advantages of cloud infrastructure, such as agility and

elasticity, they want to establish whether an open cloud architecture can

meet the applications’ rigorous throughput and latency requirements.

In particular, they need to ensure that overhead associated with

virtualization would not limit performance.

As part of the shift to the cloud, the team hopes to converge the

environment’s backbone and management infrastructures onto

10 Gigabit Ethernet. Using a single network fabric is expected to help

optimize the exibility and cost-effectiveness of the overall solution.

Traditional Architecture for Cluster Computing

The NASA Center for Climate Simulation’s research on climate change and related

phenomena, which requires extensive computer modeling, contributes to efforts such

as hurricane prediction, analysis of past weather patterns, and scientific support of

government climate policy. The cluster named Discover that has done this work for some

years uses an integrated set of supercomputing, visualization, and data-management

technologies to deliver roughly 400 teraflops of capacity:

• Compute resources: 30,000 conventional Intel® Xeon® processor cores and 64 GPUs

• Inter-node backbone: DDR and QDR InfiniBand*

• Management networking: Gigabit and 10 Gigabit Ethernet (GbE and 10GbE)

• Data store: ~4 petabyte RAID-based parallel file system (GPFS),

plus ~20 petabyte tape archive

Discover is based entirely on non-virtualized machines, so adding capacity requires

additional physical servers to be provisioned. Reducing the traditional cost and

complexity of those changes is one benet of cloud computing. Moreover, cloud

architectures add elasticity that aids in job scheduling and helps avoid operational

bottlenecks associated with long-running jobs.

CASE STUDY

Intel® Ethernet 10 Gigabit Server Adapters

Single-Root I/O Virtualization

NASA

The NASA Center for Climate

Simulation found that an open

cloud architecture using 10 Gigabit

Ethernet for both inter-node

communication and management

traffic is a viable alternative to

its purpose-built InfiniBand*-

based cluster for many large-

scale modeling applications. The

organization hopes to capture the

elasticity and flexibility benefits

of both cloud computing and

converged networking on Ethernet.

Summary of content (4 pages)

PAGE 1
case study Intel® Ethernet 10 Gigabit Server Adapters Single-Root I/O Virtualization NASA NASA’s Flexible Cloud Fabric: Moving Cluster Applications to the Cloud Network architects at the NASA Center for Climate Simulation recently began investigating the viability of running the organization’s modeling and simulation applications on cloud infrastructure, as an alternative to its purpose-built computing cluster named Discover.
PAGE 2
NASA’s Flexible Cloud Fabric: Moving Cluster Applications to the Cloud The Nebula Cloud: A Proposed Adjunct to the Existing Cluster Approach NASA’s Nebula is a cloud-based infrastructure-as-a-service environment, conceived as an alternative means of delivering compute capacity without the costly construction of additional data centers.
PAGE 3
NASA’s Flexible Cloud Fabric: Moving Cluster Applications to the Cloud The rightmost column of Table 2 shows dramatic throughput improvement in the virtualized environment when SR-IOV is utilized. In fact, the figures in this column approach those of the baremetal case, indicating that a properly configured virtualized network can deliver throughput that is roughly equivalent to that of a non-virtualized one.
PAGE 4
Conclusions 80% In terms of both throughput and latency test results using the Ohio State University MPI Benchmark, virtualized performance actually exceeded that of the bare metal (non-virtualized) case. While in testing with the Intel MKL LINPACK implementation, bare-metal servers out-performed virtualized ones, the benefit delivered by SR-IOV nevertheless demonstrates the potential for virtualized usage models.