Performance Multipliers for Data Stream Processing

“You could not step twice into the same stream; for other waters are ever flowing on to you.” – Heraclitus, Greek Philosopher (ca. 535-475 BC)

Our last posting concentrated on the definition of Data Warehouse Appliances. Today we’d like to go deeper, much deeper inside the architecture of the Netezza Performance Server^® data warehouse appliance and focus on one of the key elements of the system’s performance. We’re talking about a simple, common off-the-shelf device about the size of one’s thumbnail whose expansive capabilities at extremely low power consumption are changing the way people design high-performance, streaming systems.

The Field Programmable Gate Array (FPGA), sometimes referred to as a Programmable Logic Device (PLD), acts as a performance multiplier in the NPS^® system, increasing query processing effectiveness by a factor as high as 5X and greatly reducing the need to move superfluous data through the system.

Shining Some Light on a Key Performance Engine
Philip Howard of the Bloor Group recently wrote that Netezza has kept its “light under a bushel” regarding the advantages of FPGAs in the NPS data warehouse appliance, so we’ll try to provide a little more information here.

FPGA technology is one of the key enablers of Netezza's performance and I'll talk about that some. But I hope to show enough examples about the market for FPGAs in other domains as well, where they are providing a fundamental advantage – particularly in the domain of stream-based processing.

So, What Is an FPGA?
The OpenFPGA organization, an industry association of over 100 registered participants, defines an FPGA as, FPGA: “[An] integrated circuit able to change interconnectivity of a large number of fundamental computing components via configuration information stored in onboard static RAM”.

Off-the-shelf, an FPGA is like blank canvas for systems designers. It provides the speed and computational performance of hardware design while allowing for a very high degree of programmability and such characteristics as:

• increasing speed & density;
• increased I/O pin count and bandwidth;
• lower power;
• lower cost per gate; and
• integration of hard IP (e.g., multipliers, PowerPC cores)

We'll discuss it more in the second installation, but FPGAs are used for streaming data applications in a wide sweep of product markets – from consumer electronics, to medical applications to high-performance computing. And the technology is evolving rapidly; in many cases taking on roles that have previously been the domain of CPU technologies.

In fact, the technological trends to drive down cost and power while driving up performance are moving faster for FPGA devices than Moore's Law is doing for CPU (e.g., Intel or AMD) technology. Moore’s Law suggests a doubling of CPU performance approximately every 18-24 months, but FPGAs are progressing much faster. A paper submitted to the Association of Computer Machinery (ACM) by Keith Underwood of Sandia National Laboratories details how FPGAs are outpacing gains in CPU technology, captured best with this quote:

Underwood’s paper focused on the FPGA’s growing advantages in floating point calculation performance, but others have talked not only about processing horsepower but also advances in both I/O and memory bandwidth speeds, as shown in the following figure.

Source: “OpenFPGA BOF Presentation” at SC05 Aussie Schnore (GE Global Research) & Malachy Devlin (Nallatech), 16 Nov 2005