White Papers

ManualsBrandsDell ManualsConverged InfrastructureHigh Performance Computing Solution Resources

Dell - Internal Use - Confidential

HPC in an OpenStack Environment

Joseph Stanfield and Nishanth Dandapanthula, June 2014

Introduction

As the concept of cloud computing continues to expand its market reach, many companies have

discovered the advantage of encapsulating configuration details into a virtual setting allowing users to

build a customized environment and run it in the cloud as the need arises for computational resources.

Cloud computing would also seem to go hand-in-hand with a production HPC environment, offering

virtually unlimited storage with instantly available and scalable resources. OpenStack is the open source

cloud computing platform used in this study.

The applications in the HPC domain have massive requirements in terms of CPU, memory, I/O and

interconnect. Traditionally HPC applications have been run on physical clusters, but with the trend

moving towards cloud computing and virtualization, we wanted to see how these applications fare in a

virtualized environment. In theory, the ability to scale out available resources on per-user basis would

boost productivity and lower the total cost of ownership of the cluster. But, how does the performance

of virtual machines (VM) compare to bare metal servers (BM)?

In this blog, we’ve set out to compare the performance of a physical server with a bare metal installation

and a virtual machine, using a single node in similar environments, with identical resources. The bare

metal machine is a physical server with just a minimal OS installed. VM refers to the virtual machine

running on a hypervisor on this bare metal machine using all the cores and memory of the bare metal

system, thus both having the same configuration.

Consider a scenario with multiple project development needs, where users require a range of custom

platforms for their individual projects. There may be a need for a whole server or multiple servers for

various reasons such as application development, code beta testing, sharing a stable and uniform

platform among collaborators etc. An administrator would be able to easily deploy an environment

tailored to each user without having to re-provision the entire server farm for each project. Once the

user is done, the VM’s data or the VM itself can be archived for future use.

We study the differences in performance and the overhead raising from the use of VMS when compared

to BMs in an HPC space. We present analytical results and weigh the pros and cons of each approach.

This is the first in a series of blogs where we will evaluate virtual machines, Linux containers, and bare

metal servers and their respective tuning options from the perspective of applications in HPC domain. In

future posts, we will expand this study at scale by introducing the interconnect component.

The test bed has a head node and a compute node with bare metal installations of RedHat Enterprise

Linux 6.5. We installed RDO OpenStack on the head node and used that to add the compute node to the

resource pool. The VMs are deployed solely on the compute node. The details of the test bed and the

Summary of content (6 pages)

PAGE 1
HPC in an OpenStack Environment Joseph Stanfield and Nishanth Dandapanthula, June 2014 Introduction As the concept of cloud computing continues to expand its market reach, many companies have discovered the advantage of encapsulating configuration details into a virtual setting allowing users to build a customized environment and run it in the cloud as the need arises for computational resources.
PAGE 2
BIOS configuration used are shown in Table 1. The BIOS settings chosen for this study on the bare metal machine are typical HPC recommendations for optimal performance [310]. Table 1 Test Bed Configuration Head Node Server Memory Processor OS OpenStack Deployment Bare metal Node / VM node Server Processor Memory BIOS OS Interconnect Hypervisor Dell PowerEdge R720 16 x 8GB @ 1866 MHz 2 x Intel Xeon E5-2697 v2 @ 2.70GHz RHEL 6.5 Icehouse-3 RDO PackStack Dell PowerEdge C6220 II 2 x Intel Xeon E5-2680 v2 @ 2.
PAGE 3
WRF 3.3 MILC 7.6.1 manufacturing, crash testing, aerospace industry, automobile industry etc. Open source application which helps Conus 12KM Rating in atmospheric research and (Jobs/Day) forecasting MIMD lattice computation is an open Input file from Intel Rating source quantum chromo dynamics Corp.
PAGE 4
Figure 1 Stream Triad memory Bandwidth 120000 100000 Bandwidth in MB/s 96135 80000 78727 77527 BM-NI-enabled VM-on-NI-enabled-BM 70437.5 60000 40000 20000 0 BM-NI-disabled VM-on-NI-disabled-BM Figure 2 and Figure 3 show the performance of the applications listed in Table 1. The graphs show VM-1 and VM-2 performance relative to BM-1 performance. Most of the following applications are impacted by the CPU characteristics such as core speed, core count etc. and memory bandwidth requirements.
PAGE 5
Performance of VMs relativeto BM 1.20 1.00 1.00 0.80 1.00 1.00 0.96 1.00 0.95 0.99 0.98 1.00 0.88 0.82 0.75 0.60 0.78 0.71 0.73 0.40 0.20 0.00 LU IS FT BM-NI-disabled EP VM-on-NI-disabled-BM CG VM-on-NI-enabled-BM Figure 3 Application performance relative to BM Performance of VMs relative to BM 1.20 1.00 0.80 1.00 1.00 0.83 1.00 0.90 0.86 1.00 0.88 0.88 1.00 0.89 0.93 0.95 0.77 0.60 0.40 0.20 0.
PAGE 6
mentioned previously this is all contingent on the application’s requirements such as computation, communication (Interconnect), memory bandwidth etc. These applications are customized and tuned to reap maximum performance from the physical layer of the hardware. With cloud, performance may take a hit in terms of computational performance and I/O. Alternatively, in some scenarios, a 25 % drop in performance may be acceptable considering the flexibility and convenience cloud computing provides.