|
|
|
Ask the expert
Add capabilities to the data warehouse dynamically with Teradata's extensibility features.
|
An active user community
Empowering users of business intelligence adds ROI for your enterprise and confidence for your team.
|
Enterprise BPM
Integration of technologies from Hyperion and Teradata result in an enterprise architecture for Business Performance Management.
|
Enterprise data warehousing requirements drive the design and architecture decisions of the Teradata platform.
|
Tech support
Need technical help? A Teradata Certified Master answers your queries.
|
|
|
|
|
|
Teradata platform: Built for a purpose-driven data warehouse
Keep up with the best new technology without skipping a beat.
by Robin Jessani
and Keith Rimmereid
Computer platform engineering is as much about the art of effective trade-offs as it is about the science of design. Trade-offs are imposed by a multitude of technical and business
factors and an ever-changing mix. Technical factors include architectural design, component performance,
compatibility, and size and heat characteristics. Business factors include cost, time to market and the competitive landscape.
The computer industry has matured to the point where the repertoire of technology options available to platform design teams is quite uniform. The differences in the resulting systems stem more from the design trade-offs than from differences in the base technology. However, in the Teradata platform, it's enterprise data warehousing requirements that drive the design
and architecture decisions.
The NCR Massively Parallel Processing (MPP) platform is designed for the sole purpose of running the Teradata Database as efficiently as possible in support of data warehousing. Let's look at five key design criteria that drive development decisions behind the NCR MPP platform family:
 | Scalability: Uniformly increasing capacity and performance to meet business needs |
| Availability: Eliminating or minimizing the impact of component or module failures |
| Manageability: Minimizing the human effort necessary to run
the system |
|
High-performance technology: Rapidly adopting and exploiting
the highest-performing technology |
| Growth with investment protection: Expanding and evolving the system while avoiding the financial penalty
of obsolescence |
The "purpose-built" approach of the NCR MPP platform family has been refined based on customer installations throughout the technology's evolution. The goal of the purpose-built approach is to let customers focus on business, not on managing the technology—thereby reducing time to solution and increasing ROI.
Scalability
Teradata Warehouses often start small and grow rapidly and unpredictably for several years. Over a four-year period,
a mid-sized North American retailer grew its Teradata Warehouse by a performance and capacity factor of 11 by expanding the platform—not replacing it. Starting with less than 300GB of data, a single subject area and a handful of users, the data warehouse grew to over 3TB of data, many subject areas and hundreds of users. A key enabler of this growth was the NCR MPP platform's ability to expand in a predictable, linear manner consistent with the increased workload needed by the business.
It is the shared-nothing architecture
of the NCR MPP platform that supports scalable growth to nearly any dimension. Shared-nothing architecture is a method of building a platform that eliminates resource bottlenecks resulting from shared resources. An example of a shared resource is the internal memory to CPU bus, to which CPUs are connected. Only one CPU can use the bus at any one time, so the more CPUs on the bus, the more time each spends waiting for access to the bus and not doing useful work.
Today's NCR MPP platforms are built upon two CPU node building blocks, thus minimizing bus contention (see figure 1). The two CPU node building blocks are interconnected via a scalable interconnect which is used only for database messages and not for memory or disk access. Each node is provided
direct, dedicated access paths to disk storage, eliminating another point of sharing contention.
|
Figure 1
|
|
Comparison of a large Symmetric Multiprocessing (SMP)/Non-memory Uniform architecture (NUMA) to the Teradata shared-nothing Massively Parallel Processing (MPP) architecture.
|
In contrast, most of today's systems
are built using a shared memory and shared disk architecture. These systems are general-purpose in nature, and one
of their primary responsibilities is to be compatible with the huge base of non-data warehousing applications, which have been written for the shared-resource model. These shared-resource architectures suffer from the performance and scaling loss of shared memories, disks and system interconnects, when used in
a data warehouse environment.
Availability
Platform architecture and design insulates the database and applications from many types of hardware component and module failures. Modular packaging isolates the failures to specific components, which can be quickly and easily replaced by Teradata Customer Services, minimizing system impact.
All NCR MPP platform configurations by default include appropriate redundant components and unique, integrated, automatic failover capabilities to ensure a high level of out-of-the-box availability.
The NCR MPP platform contains redundant hardware components so that a failure of a redundant component does not impact the operation of the database or the end-user experience. An example is the pair of quad fiber channel disk array adapters that are contained in each node. In normal operation, the node disk I/O workload is balanced across ports on both adapters. In the event of an adapter failure, the surviving adapter automatically takes over the full workload with
no impact to database availability or
performance. The system management subsystem alerts the Teradata Global Support Center to initiate appropriate repair action.
A unique built-in attribute of the Teradata Database is the ability to migrate workload from a failed compute node to surviving nodes (see figure 2). This automatic failover capability is supported by the NCR MPP platform via the "clique" architecture. Cliques are sets of nodes connected to a storage subsystem with enabling infrastructure to allow workload failover between and among the nodes of the clique. The clique
configuration relies on the large I/O
connectivity between the disk arrays
and the nodes, utilizing between four and eight fiber channel connections per node. The high I/O bandwidth provided by the platform enhances performance for the data intensive data warehouse workload and also improves availability.
|
Figure 2
|
|
In a failure situation, the software will reconfigure itself through a two-minute restart process and, as evenly as possible, redistribute the AMPs from a failed node to the surviving nodes within a clique.
|
Some Teradata customers have
their systems configured with four-node cliques. In the event of a failed node, three nodes carry the workload
of four, thereby reducing performance
by more than 30%. To minimize the performance impact of a failed node, there is an option called Large Cliques that increases the number of nodes in a clique up to eight, reducing the performance impact of a node failure to
as little as 15%. To eliminate the performance impact of a failed node, a Hot Standby Node option allows the entire workload to be migrated to a node in waiting. Redundant, integrated fiber channel switches are the enabling hardware support for the Large Clique and Hot Standby Node capabilities.
For additional information on Teradata availability features, see "Data warehouse availability" by Todd Walter, Teradata Magazine, Vol. 5, No. 2, or visit the Teradata Magazine archives. For information on business continuity solutions, see Services.
Manageability
To simplify platform command, control and monitoring, a single operational
view is provided by an integrated systems management infrastructure. Many of the NCR MPP design decisions reflect the ease of management philosophy for Teradata Warehouses.
To support integrated systems management, redundant management computers are embedded within the platform cabinet structure. The computers are each connected to a dedicated, redundant, parallel management system network. Intelligent software agents communicate to all of
the components and record all of the events that the components report.
The integrated system management infrastructure monitors and controls the system, performs routine events such as orderly start up and shut down, and protects the system from disruptive failure such as a power loss or extreme heat. Like the system itself, the high-performance parallel management infrastructure scales with system size and easily accommodates the thousands of components and subsystems typically found in large
system configurations.
The basic management approach is to collect the individual events that various components report, put the events into a context that intelligently interprets system status and present the appropriate actions or information messages to the system management console. Information may also be forwarded to an enterprise management system or the Teradata remote support facility. The software will collect and retain information to enable proactive detection of significant events and can initiate actions to correct a pending failure.
For example, suppose a disk controller was reporting an abnormal amount of retries. An individual retry itself is masked and handled automatically by
the disk array, but if the number of
retries hits a threshold, that event will
be automatically reported and can initiate a proactive maintenance action for replacement. This type of event would be retained in a permanent database at the remote support center and compared against historical aspects of the system against norms that are occurring across all of the systems within the Teradata customer base. The information that customer support collects and
analyzes and the intelligence within
the software allow proactive support intervention, which increases the
reliability and the availability of the
customer's data warehouse.
High-performance technology
One of the challenges in system design
is to incorporate the ability to adopt new technology as it becomes available without a significant re-engineering effort that might affect customer cost or delay the benefits of new technology.
The NCR MPP systems use modular design and architecture (see figure 3), with specific system functions in discrete packages that can be independently evolved without significant impact to
the rest of the system. This approach has benefited Teradata customers, enabling them to leverage the latest technology without being subjected to the disruption of technology churn.
|
Figure 3
|
|
Teradata's platform strategy is to invest in NCR value-added technologies for increased system availability and performance and leverage best-in-class components from industry leaders to deliver a complete, optimized and integrated system purpose built for data warehousing.
|
The Teradata platform strategy is
to leverage best-in-class components from industry leaders and select specific areas, adding significant value to the Teradata Warehouse. Harvesting industry technology, developing critical value-add technology and owning the entire environment through the integration
of all of the components are key to
delivering an optimized platform for
the Teradata Database.
Teradata engineering performs
pre-selection, pre-testing and complete certification and delivery to customers, eliminating significant effort, risk
and cost in the customer environment
by doing all this work within the
engineering process.
Let's take a look at one of the areas where there is substantial value added, specifically to the Teradata Database.
The BYNET interconnect was designed
to specifically enhance the ability to perform parallel database operations. The BYNET differs from general-purpose interconnects in several important ways. It scales proportionately with the size
of the system, never becoming a performance bottleneck.
The BYNET is a redundant, fault-tolerant interconnect infrastructure, which is self-healing and auto-configuring. It provides guaranteed message
delivery, freeing the database from performing
this check. It provides semaphore facilities
to coordinate parallel database steps, thus reducing the cycles required in the Teradata Database. The BYNET also provides both point-to-point and broadcast messages, giving the Teradata Database optimizer more efficient ways to process a request. The BYNET performs in parallel the final merge/sort of the answer set, thus eliminating a serial operation and improving performance. With its
unique value-add capabilities, the BYNET enables the Teradata Database to perform better.
The most visible area that harvests best-in-class industry technology is
the node CPU. Teradata has a long-standing engineering relationship with Intel and leverages the industry-leading Intel CPU technology for the system compute nodes. The NCR MPP systems are designed to evolve and take advantage of the evolution and investment that Intel is making in moving the CPU
technology forward.
The NCR 5400 MPP Server is the
latest in a long line of server platforms that have evolved with the Intel processor technology. The NCR 5400 is faster and more reliable, it requires a smaller system footprint than its predecessors and it
can contain up to 10 nodes within a single cabinet.
It is linked with the high-performance NCR Enterprise Storage subsystem, designed to be modular and to map to the Teradata architecture. It is optimized for the very high I/O environment of the Teradata Database and is integrated and managed through the system's management single operational view.
Growth balanced with
investment protection
The ability of the NCR MPP server to easily adopt new technology improvements has led to a consistent release of
a new generation server every year. The platform and the Teradata Database have been engineered with a capability called coexistence to support multiple generations of servers within a single system.
By enabling expansion through coexistence, customers have the option
of expanding their systems to include
the latest server offering, while at the same time reaping a return on the initial technology investment and also taking advantage of the price performance curve as it evolves.
For example, assume an original system with the NCR 5350 Server and Teradata Database with seven parallel units (Teradata AMPs) running per node. As business needs grow, the system can be expanded and obtain at least 60% more power by using the latest technology of the NCR 5400 Server.
To utilize coexistence, the workload will be balanced proportionately to
the power and performance capabilities of the node and disk combination. To balance the systems, we'd recommend putting 60% more units of parallelism (AMPs) on the faster node; in this example, it would be 11 AMPS per 5400 node. We'd also configure 60% more disk storage capacity and equivalent I/O capacity with the node. Future expansions would follow the same scenario.
Over the years, hundreds of coexistence systems have been deployed by customers, helping to refine and broaden this capability to where it now is certified for up to five generations of servers.
Conclusion
Teradata engineers and integrates the entire system and relieves the customer of many integration risks and activities. The system is pre-staged and tested in
the customer-specific configuration and is delivered with the operating system and database software installed and appropriately configured.
The Teradata install team has only a modest amount of reassembly and final testing to perform before turning the system over to the customer. Thanks
to the integrated process, Teradata
customers typically can begin loading data and running queries within days
of initial delivery—and begin reaping business value.
The Teradata "built for data warehousing" approach eliminates the unexpected, reduces risk and allows customers to focus on driving the highest return on their investments, not only today but into the future. T
Robin Jessani joined Teradata in 1992. She has served in many roles surrounding the Teradata Warehouse and is currently a Teradata Warehouse Solutions marketing specialist.
Keith Rimmereid is a senior Teradata consultant with Teradata Data Warehousing Sales Support in San Diego. Keith works closely with many key Teradata customers to help them understand the Teradata technology and how to best apply the capabilities to meet their business needs.
Beauty that's more than skin deep
|
The cabinet of the NCR 5400 Server
is more than a pretty shell. Its ingenious design actually optimizes performance and reliability.
by Eric Schoeniger
It's hard to think outside the box when your job is designing boxes. But that's exactly what Teradata's engineering team did when it developed an innovative cabinet for the NCR 5400 Server. The result is a patent-pending design that optimizes performance, reduces space requirements, improves serviceability and enhances reliability.
Built for large-scale data warehousing, the NCR 5400 boosts Teradata Warehouse performance by 20%. What's more, EM64T technology in the server nodes provides 64-bit capabilities. "We wanted the NCR
5400 to be recognizable as a Teradata server,"
says Imad Birouty, senior marketing manager for Teradata. "But we also wanted to show that it's a
64-bit transition platform."
Form and function
So the engineering team got busy on what proved to
be a multi-year design process. While the Teradata team focused on the cabinet's technical specifications,
it brought in industrial design firm DD Studio for the
exterior look and feel.
Birouty's team provided DD Studio with requirements such as dimensions, strength, operational use, shipping practices and construction quality. It also explained that the cabinet needed to reflect Teradata's core traits of enterprise quality, high performance and value.
DD Studio analyzed the designs of other leading servers to identify the state of the art and be sure the new cabinet would stand out. The designers focused on function, form, materials, color and "language," or the design details that create an overall impression. "The cabinet needed a distinct personality," says Sandy Stewart, the feature realization team leader for the NCR 5400 and a senior hardware engineer for Teradata. "It needed to communicate speed,
quality and reliability."
DD Studio came back with 15 concepts in five families such as "digital/techno," "wave/group" and "stealth/superhero." The Teradata team narrowed that to four choices. "Each had pros and cons," Stewart says. "For example, we liked the waves. But in an installation with a lot of cabinets, it would have made you dizzy."
The favorite concept came from the "directional/car styling" family, meant to suggest the speed and complex form of a sports car. The Teradata design team then asked DD Studio to integrate additional elements to create a truly winning design.
Solar plexus
Teradata engineers, meanwhile, were moving forward with the cabinet's interior. "The interior layout has to
be visually appealing. There can't be a mishmash of colors and textures and materials," Birouty says. "But even more important, it needs to be functional
and serviceable."
That was a challenge, because the team had to squeeze 10 dual-processor Intel notes, five uninterruptible power supplies, a BYNET interconnect and various other components into a space of just 48 cubic feet (77 inches high, 24 inches wide and 45 inches deep). Those dimensions allow the server to ship by typical "freight" methods and fit in a standard elevator.
The cabinet's weight had to be evenly distributed, with heavier components near the bottom. Casters had to be sturdy and strategically placed to support 1,500 pounds of technology. And all components had to be accessible for easy service—including the 70-pound server nodes. That meant designing the topmost racks to slide out and tilt down for easy access.
Finally, cabling had to be cut to exact lengths and carefully routed to fit in the tight space. "We spent many hours doing mockups to validate cable routing and serviceability," Stewart says.
Feeling the heat
There was one problem left. A high-end server runs hot. The DD Studio design called for a large grill in
the cabinet door to maximize airflow. Trouble is, air wouldn't be evenly distributed to all components. Tests showed that the bottom server nodes would draw in cool air, while the top nodes would get warm air. Worse, hot air vented out the back would be drawn back into the top of the grill. "We couldn't cool more than six nodes," Birouty says. "So we had to use the cabinet itself to improve airflow."
Teradata engineers came up with an inventive solution. They replaced the large grill with a smaller grill near the floor. Then they added depth to the cabinet door, sealing the edges and creating an airflow channel.
Now, cool air is drawn into the bottom of the door and channeled up the front of the cabinet. The cool air is equally distributed to all nodes, and the hot exhaust isn't drawn back in.
By supporting up to 10 server nodes in a single cabinet, the NCR 5400 significantly boosts performance. It also delivers a 40% smaller footprint, optimizing data center space. The thoughtful placement of components and cabling improves serviceability. And with the patent-pending approach to cooling, storage reliability is increased by an impressive 30%. Beautiful!
|
© Teradata Magazine-March 2006
back to top
|
Quick Links
Print
Send to colleague
PDF
E-mail us
