OpenStack Case Study: iSuperGrid and Mirantis

The Challenge

Cloud Computing and high performance computing (HPC) workloads might seem a natural fit. They do share key attributes:

• Require a versatile infrastructure, where compute cycles can be brought to bear in flexible fashion on very diverse workloads.
• Rely on fluid management of resource configurations.
• Are well-suited to distributed computing problems, where processing can be spread over many units of physical hardware.

But in practice, HPC workloads impose unique demands on the cloud: provisioning, configuration of compute resources and storage resources, and lifecycle management have needs of their own. In fact, running HPC workloads on conventional cloud infrastructure can lead to performance slowdowns, higher costs, and obstacles to the rapid iteration of workload cycles that HPC requires.

Information SuperGrid Technology (ISGT or iSuperGrid) is a cloud service provider whose system and application software make it possible to integrate distributed and parallel grid supercomputers into a massive and seamless global cloud computing platform. In this platform, each computer node leverages and harnesses the power of the other nodes to scale up processing capacity and capability when needed.

iSuperGrid has built an innovative new model to deliver this cloud platform to organizations with HPC needs. To marry the flexibility of cloud computing with the complexities of HPC, they turned to the OpenStack cloud infrastructure technology.

‘Supergrid’ Cloud

Working with Mirantis, a company that provides engineering services for OpenStack technology projects, iSuperGrid is building out a distributed OpenStack cloud network, numbering into the thousands of nodes, providing HPC and versatile hosting to a variety of emerging markets.

As part of its business model, iSuperGrid offers a fixed-size, pre-certified cloud server rack set, deployed as an on-demand resource for HPC workloads to a broad range of government, academic and commercial organizations. Each organization’s OpenStack cloud rack is, in turn, networked into a larger, collective ‘supergrid’. This way, any iSuperGrid customer can then offer spare HPC-capable compute cycles during off-peak hours, such as nights and weekends, to run HPC workloads to third parties. Configurations are changed on the fly to fit different types of HPC workloads, such as those that require large memory resources but little persistent storage, or vice-versa. These cycles can be sold directly, or brokered via iSuperGrid to anyone needing fixed-term cloud services.

OpenStack Use Case

HPC Workloads
Because it specializes in HPC, iSuperGrid delivers cloud services to health, education, industrial, logistics, mineral extraction, media/entertainment, and other markets in China and worldwide. What these diverse compute workloads have in common is their appetite for specialized, calculation-intensive compute operations.

Examples include:

‘Smart utilities’: municipal services such as electricity distribution, energy consumption, real-time on-road traffic analysis Science and engineering: Mineral and geology analysis, weather forecasting, biomedical research, chip and instrument designs

Media: Digital content rendering, delivery, and distribution

Logistics: batch and real-time calculations of equipment deployment and utilization, routing optimization, and throughput

The OpenStack Strategy
The iSuperGrid business strategy calls for delivering these services with dynamic provisioning that scales in a pay-as-you-go offering. Making it cost effective relies on high-efficiency utilization and rapid turnaround of allocation, so that compute supply and demand move in tight synchronization. OpenStack’s flexibility and scalability makes it a good choice. In order to base this solution on the OpenStack technology, iSuperGrid asked Mirantis to provide architectural and implementation guidance.

Technical Requirements

Performance Appetite
Because HPC workloads are by definition compute intensive, Mirantis first turned to benchmarking compute performance. In conventional cloud environments, virtualization throttles back the underlying hardware compute horsepower available to applications, by allocating machine cycles across multiple virtual machines. An HPC performance test showed the scope of the challenge. Comparison of a deep analytics workload on Amazon Web Services (AWS) vs. a dedicated HPC cluster on commodity hardware found the AWS cloud ran the workload 30x slower than the dedicated ‘non-cloud’ servers. Given that cloud environments such as AWS are optimized for deploying Web and transaction applications, it should come as no surprise that they perform sub-optimally. The question was then, could OpenStack do better?

OpenStack and Cloud Provisioning
Direct access to specialized low-level machine resources is essential to making it work in the cloud. But it’s more than raw resource access: rapid, cost-effective provisioning of incremental compute cycles with minimum overhead must feed bursts in processing appetite demanding quick access to raw compute power. Workloads such as smart-cities electric grid consumption tracking, real-time-analytics, simulations, logistics tracking, mineral extraction analysis, and similar performance-intensive applications cannot be satisfied by squeezing more work out of a pre-configured, fixed server footprint.

The OpenStack Solution

Cloud Configuration
The first iSuperGrid cloud, engineered and deployed on OpenStack in collaboration with Mirantis, featured approximately 50 server systems configured in three geographically dispersed datacenters across China. It offers both database and object storage, plus Storage-as-a-Service to support elastic compute node storage. These key cloud services are also configured for high availability.

OpenStack HPC Architecture
Working over a period of eight weeks with a team of six globally distributed engineers, Mirantis designed an architecture that would use OpenStack Essex components to blend the compute properties of a dedicated HPC cluster with the flexibility of a cloud deployment.

Bare Metal Provisioning
A key element of this architecture was bare-metal provisioning ‒ deploying well-defined, fully-functional OpenStack images on unconfigured hardware. Working from a boot server to connect to unprovisioned servers, the OpenStack cloud controller constructs a kernel that bootstraps a fully functional OpenStack OS image and connects it to the cloud fabric. The architecture designed and implemented by Mirantis enables rapid linear expansion or reconfiguration of the total available compute cycles. In this way, the iSuperGrid OpenStack cloud adds fresh hardware transparently, increasing its overall capacity with minimal latency or disruption.

Full Scale Supergrid
With the successful conclusion of the pilot, iSuperGrid is now rolling out a full-scale cloud environment with several hundred servers, on both sides of the Pacific Ocean. By using OpenStack to transparently expose capabilities unique to a particular server configuration directly to HPC workloads, without sacrificing the resource elasticity of the cloud environment, iSuperGrid has a highly fluid, flexible environment, at virtually unlimited scale. What’s more, it delivers the environment using an innovative business model to deliver raw compute power both to workloads needing sustained utilization as well as to workloads with on/off-peak processing needs.

About the Author

Boris is co-founder and executive vice president at Mirantis, a leader in engineering services for OpenStack Cloud, He is an active board member of several IT companies, and chairman of AGroup, a leading human resource management and payroll solutions company in 22 countries. In addition, he is involved deeply in the investor community.

In the past, Boris served as vice president of business development at R&K, one of the largest IT conglomerates in Russia. He was also founder and CEO of Selectosa Systems, a software outsourcing company that was acquired in 2006.