Whether it’s a pharmaceutical company analyzing clinical trial data, a financial services institution looking at real-time changes in market conditions, or a manufacturing firm investigating supply chain issues while monitoring customer concerns on social networks, organizations today must perform multi-step analyses on large volumes of diverse datasets to derive actionable information upon which to make critical decisions.
To perform these multi-step analysis routines, organizations need new high performance computing (HPC) capabilities. And it’s not your father’s HPC. Increasingly, what is being used is a more commercially-oriented solution, which requires an enterprise-grade infrastructure. Additionally, while in the past it was common to select systems on their ability to perform one task well, now the entire HPC infrastructure used in these computational processes must be optimized.
Basically, big data analytics today comes down to optimizing workflows. And this should drive organizations to examine their HPC infrastructures.
Just as triathletes might spend great amounts of time and energy improving their running, swimming, and bicycling, these efforts will not pay off if the transitions are bungled. Similarly, with big data analytics workflows, an organization certainly should seek to accelerate each step in the process while making optimal use of resources. But now the intermediate parts of the process, as well as the entire end-to-end chain of computations, must also be taken into account to ensure workflows are speedy, sustained, and economical to perform.
New demands drive need for change
In today’s world, speed is the name of the game. Whether analyzing market data to make critical business decisions, or running data-intensive simulations to better understand physical phenomenon, analytic processes must be carried out in ever-shorter time spans to be of value to an organization.
To accelerate analytic workflows requires an appreciation of the changing nature of big data. In the past, computational analysis had to deal with the growing volumes of data. That’s still the case now, but in today’s world of big data, there are two other dimensions to consider: data variety and data velocity.
From the variety perspective, more of the data used in organizations today comes in different formats and from multiple sources including intelligent sensors, GPS-enabled smartphones, video feeds, and social media sites. Working with such unstructured data can require additional computational steps within an analytics workflow. For example, raw data from these sources might need to be extracted, converted, and run through one algorithm to be useful in the next step of an analysis.
From a velocity perspective, more data today is collected from real-time streams, has a shelf life, and needs rapid analysis to be of use. For example, social networking content can change rapidly. Glowing praise for a product one day is quickly replaced by scorn the next when a problem arises. The implication here is that to get the best value from the wealth of social media information or intelligent sensors, the data mining and analysis must be done in real time.
HPC infrastructure challenges
This combination of big data and an increased reliance on multi-step workflows introduces new infrastructure challenges.
In the past, large volumes of data could be pre-loaded or pre-staged for processing and analysis. The idea being to have data in place to satiate a system’s CPUs. This does not work in today’s analytic workflows. Typically, results from each step feed the next, thus negating the possibility of pre-loading data.
As a result, at each step of a process data must selectively be staged on storage systems with appropriate I/O and throughout characteristics and moved over a network infrastructure at the right time and right speed to ensure optimal use of CPUs. For example, an analysis might require the capturing of an image from a video. That image might then need to be rendered using another algorithm to put it into the correct format. And then the resultant file might need to run through an image analysis routine to extract the desired information.
The challenge here is that it is not just a matter of matching storage volume to a routine. Organizations must also be concerned about the management of diverse storage needs throughout the entire workflow and processing life cycle.
For example, each step in an analytics workflow can produce very large volumes of intermediate data. A suitable HPC infrastructure must be capable of moving, managing, and saving this data for use in check point restarts if computations are disrupted.
Furthermore, addressing the computational, storage, and networking performance and capacity issues of today’s big data analytics workflows must be done with an eye on power consumption. Simply adding systems to boost performance or meet capacity needs can push power and cooling requirements to a level that cannot be met in an already stressed data center.
Taking these factors into account, organizations can no longer pick an HPC solution based on a single benchmark, such as a server’s maximum processing power or its ability to run a particular database test faster than a competitor’s solution. Instead, organizations need to examine the total workflow and its associated data movement. Specifically, organizations must examine the various tasks in their big data analytics workflows and match the requirements with suitable HPC solutions. And as organizations increasingly rely on these workflows to make critical business decisions, HPC will become more mainstream. This in turn will drive the need for enterprise-grade systems.
IBM as your technology partner
To meet the infrastructure needs of today’s big data analytic workflows, IBM offers a broad range of systems of varying price and performance.
The IBM big data analytics hardware portfolio includes IBM System Storage® such as the DCS3700 High Performance Storage in addition to IBM System x®, POWER7®, and Blue Gene® server families. HPC clusters based on IBM System x systems are x86-based servers that make use of the most current multi-core processor technology. The servers are available in tower, rack, or blade formats.
Servers in the System x family include:
- iDataPlex® which is designed for applications used in CAE, risk analytics, and energy exploration. These servers are ideal for use in growing data centers that need additional performance but have space, power, and cooling constraints.
- Intelligent Cluster™, which is an integrated solution that offers significant price/performance advantages for many high-performance workloads by harnessing the advantages of highly innovative servers. Intelligent Cluster integrated HPC solutions include servers, storage, and interconnects that are factory-integrated, fully tested, and ready to plug into a data center. These solutions are ideal in supporting the work of a department or line of business.
- IBM eX5, which is designed for applications that have large memory requirements and need high performance. These servers are well-suited for applications that use and manipulate large datasets. They offer memory scaling and support for Solid State Drives (SSDs) for improve performance.
The IBM cluster solutions integrate all components, including switching, storage, job-scheduling software, and a message-passing interface. Additionally, because of its expertise addressing big data analytics issues, IBM can help optimize many application environments on a given cluster.
Organizations that need higher levels of computational muscle can turn to IBM POWER® systems. New systems based on the POWER7® processor offer the sustained application performance needed to meet today’s big data analytics challenges. The systems support very large SMP memory, fast interconnects, and very dense packing. The higher end models are routinely used by national laboratories to conduct critical research; systems based on the POWER7 Blade increasingly are being used in more mainstream HPC applications such as handling workloads in commercial organizations. Further, when ultra scalability and unprecedented energy efficiency is important, IBM System Blue Gene® should be considered.
Beyond the System x and POWER server systems, IBM’s technology expertise comes into play in other ways, too.
The IBM General Parallel File System (GPFS) helps accelerate parallel calculations on very large data sets. GPFS improves performance by reading and writing data in parallel across multiple disks or servers. Because the data is stored across a pool of disks or servers, GPFS enables a higher level of fault resilience for clusters and grids across multiple locations. These capabilities can vastly improve big data analytics workflow performance.
When it comes to handling the greater volume of real-time data sources, IBM InfoSphere® Streams helps organizations rapidly ingest, analyze, and correlate information as it arrives from thousands of real-time sources. In particular, InfoSphere Streams can be used to perform complex analytics of heterogeneous data types including text, images, audio, voice, VoIP, video, police scanners, web traffic, email, GPS data, financial transaction data, satellite data, sensors, and any other type of digital information relevant to an organization. And IBM InfoSphere Streams lets organizations quickly develop new applications to meet shifting priorities.
In the near future, organizations will be able to benefit from Watson-like technology, making use of application and machine learning to bring a more expansive variety of data into analytic workflows. There is already some movement on this front in healthcare. IBM has partnered with Nuance Communications, which offers a speech recognition and Clinical Language Understanding package. And this month, IBM partnered with WellPoint, who will develop and launch Watson-based solutions to help improve patient care through the delivery of up-to-date, evidence-based health care.
Beyond infrastructure hardware and software, IBM offers organizations other help. IBM has expertise in most, if not all, big data analytics areas. IBM’s knowledge of the applications in this domain allows the company to optimize solutions for particular industries. Additionally, IBM has research centers around the world that routinely team with organizations and government agencies to tackle First-of-a-Kind projects that integrate big data analytics efforts with IBM expertise and solutions.
For more information about IBM big data analytics efforts designed to meet the organizations today, attend the IBM High Performance Computing Virtual Event or visit IBM High Performance Technical Computing team.