To identify opportunities for using deep learning (DL) and IBM PowerAI, this chapter describes three different use cases that are based on our field experiences. For each scenario, we describe the requirement, solution, and benefits.

This chapter contains the following topics:

In many modern industries, companies require and depend on real-time responses from their applications and systems to maintain a successful business. Such customers are receptive to new technology (such as DL) and have expectations about any new technology that quickly contributes to satisfying their business goals and requirements.

When users and data scientists initially experiment with implementation, they do not consider the infrastructure that hosts their DL frameworks because their focus on the environment is the accuracy results that their models produce. However, as their familiarity grows and they implement more complex models with larger data sets, it is inevitable that the training phase takes longer. As a result, their focus shifts to the infrastructure under the frameworks. They start to evaluate differences in performance depending on the servers’ configurations. As a criteria for vendor selection, customers tend to evaluate the performance of one specific framework, and testing a specific data set such as an imagenet.

Here are examples of customer requests:

The customer requirements that are described in 7.1.1, “Customer requirements” on page 232 can be satisfied by implementing the following environment:

As a result of the proposed solution that is described in 7.1.2, “IBM solution” on page 232, there are a few benefits:

Figure 7-2 illustrates the image of the installation for IBM PowerAI Release 4 and IBM PowerAI Vision on a bare metal server.

Figure 7-3 shows the image of the installation for IBM PowerAI Enterprise with all the benefits from IBM PowerAI Release 5 on bare metal server due to the new option of Red Hat Enterprise Linux with Release 5.

Cloud computing is the standard platform for many companies from an IT consumer perspective as a viable alternative to hosting internal environments. Alternatively, from the viewpoint of IT providers, platforms must be on-premises servers and maintain the highest level of system and data security. Regarding data protection, healthcare institutions and research institutions are frequently mandated to host data within their institutions or within specific countries because of the required treatment of personal, financial, or regulatory information.

It is a common requirement to securely use servers for multiple and distinct customers. This section introduces the multitenancy environment.

Here is one of the examples of running NVIDIA-Docker on Power S822LC for High Performance Computing and IBM PowerAI on a Docker container. The purpose for using this NVIDIA-Docker solution generally is for cloud services, and internal use across the departments within the company.

Here are examples of customer requests:

In this situation, the flexibility of the platform is highly requested. Because of the GPU-enabled Docker technology, that is, NVIDIA-Docker, IBM can provide the following components in the solution:

NVIDIA-Docker with a guest operating system image: Ubuntu and CentOS

From the solution that is proposed in 7.2.2, “IBM solution” on page 236, the following benefits can be achieved:

Figure 7-5 shows the architectural diagram of NVIDIA-Docker that is installed on Red Hat Enterprise Linux to obtain the multitenancy environment for business use.

IBM Power Systems servers have been used within high-performance computing (HPC) environments since the days of the original IBM RS/6000® SP clusters. The scalability and flexibility of the hardware platform makes it an attractive choice for the HPC community. This relationship continues with the advent and popularity of DL. The usage of GPU-enabled
IBM Power Systems for HPC environments is increasing. These servers provide an alternative model compared to traditional HPC environments, where it was expected to require the largest server models with the most quantity of resources, now the GPU-enabled models provide the same processing power in a much smaller footprint. Where previously rooms were required, now we can talk in quantities of racks.

Presently, there are many data scientists conducting DL frameworks in such clusters. The largest announced environment at the time of writing is the one that will be adopted for the Collaboration of Oak Ridge, Argonne, and Livermore (CORAL) project for the US Department of Energy. This environment will be constructed by using the recently announced POWER9 based Power AC922 machines. Especially for the DL frameworks in HPC, users manage the environment with job scheduling software, such as IBM Spectrum LSF®, and recently with Spark technology to enable distributed parallel processing.

Here are examples of customer requests:

To satisfy the customer requirements that are shown in 7.3.1, “Customer requirements” on page 238, IBM provides experiences and solutions that can be integrated with other software. As a basic proposal (Figure 7-6), IBM can provide a solution with the following components:

By providing an IBM PowerAI end-to-end solution consisting of all of the components that are shown in 7.3.2, “IBM solution” on page 238, customers can obtain these benefits:

In the previous sections, we described the key components of the IBM PowerAI Deep Learning framework, summarized the benefits of the GPU-accelerated Power Systems hardware, and provided example installations and implementations of IBM PowerAI versions 4 and 5.

IBM PowerAI makes the field of DL available and consumable to every audience, including the causal user, small businesses, and traditional HPC organizations. The significance of the acceleration that is provided by GPUs results is realized in savings in both processing time and physical footprint. The smaller use cases that use capacity from a cloud provider (see IBM and NVIDIA Team Up on World’s Fastest Deep Learning Enterprise Solution) provide a low-cost route for evaluation and implementation, and in the larger cases, the number of physical servers are reduced compared to traditional computing requirements from previous decades.

If you are unfamiliar with the field of DL these example use cases provide context and perspective. Understanding what data you have is key to using DL. What connections, trends, or observations might be hidden from your existing view? Using DL for data processing provides many previously unavailable opportunities and areas of interest for academic, scientific, industrial, retail, and commercial businesses. DL has been growing in popularity for some time, but it is the current ease of adoption that is making it commonplace in many areas of our everyday life.