Remember me

Register  |   Lost password?

The Trading Mesh

Successfully deploying FPGA-based Network Adapter Cards to Deliver Ultra-low Latency Trading

Mon, 25 Jun 2012 08:00:00 GMT           

By Stephane Hauradou, VP - Accelize (www.accelize.com

 

In recent years, FPGA technology is being rapidly recognized as a game changer in ultra-low latency electronic trading, sparking the need for off-the-shelf, programmable, FPGA-based trading solutions coupled with accessible programming methodologies and design tools. Even though a number of technology providers have emerged with a variety of FPGA-based solutions tailored for the financial services industry, the reality is that only a fraction of electronic trading solutions successfully leverage FPGA technology. This is due, in part, to the very different set of skills required to understand and program FPGAs which is a definite departure from the traditional software-centric approach inherent to the financial industry. In this article we will explore some of the keys to successfully deploying ultra-low latency trading solutions based on today’s cutting-edge FPGA technology. 

 

An important first step - Selecting the right FPGA adapter card

There are several critical factors to selecting the right adaptor card for a given financial application. These factors include the FPGA technology, the need for Ethernet connectivity, the required form factor of the card, memory requirements, power supply considerations and ease of FPGA programming. The correct card will have an optimal mix of these factors for the application.

FPGA Technology: Latest generation translates tobetter performance.

A card that uses the latest generation FPGA ensures that algorithms and logic can run faster, which usually translates into lower latency – possibly THE most important factor in FPGA-based acceleration solutions. Additionally, in most cases the latest generation FPGAs have a better cost-to-size ratio which provides more real estate within the FPGA, allowing for more complex algorithms and the ability to perform more processing in parallel.  Communication interfaces are also typically faster in newer technology FPGAs, which generally correlates to lower latency. For example, the Altera Stratix® V and Xilinx Virtex®-7 FPGAs support Gen3 PCI Express allowing transfer speeds up to 7GB/sec to the server memory (in x8 configuration), whereas the previous generation of Altera and Xilinx FPGAs only supported Gen2 PCI Express and transfer speeds under 4GB/sec . Finally, latest generation FPGAs often support faster memories, improving I/O latency.

Ethernet connectivity: Creating flexibility and scalability

Market data from leading exchanges is increasingly broadcast using 10G Ethernet and some exchanges are even transitioning to 40G Ethernet. Choosing an FPGA card that offers multiple 10G links and a direct connection to the FPGA have a clear advantage both in terms of scalability and latency. Coupled with optimized 10G Ethernet MAC and UDP/TCP hardware stacks within the FPGA typically translates into wire-to-FPGA fabric latency under 200 ns for raw market data.

Card form factor: Ensuring fit and function within the server

A large percentage of trading appliances utilize blade servers, particularly in co-located environments. These servers typically have a 1U form factor, most of which only accommodate 1U adapter cards within the box. A low-profile, half-height FPGA adapter card has a definite advantage over a conventional height card in that it will fit within the server and allow adequate airflow for cooling.

Memory: Using onboard memory to increase processing performance

Trading appliances can always use more onboard memory, in particular fast and high density bulk memory such as DDR2 or DDR3 SDRAM which is required for building large order books, and fast access, low-latency memory such as QDRII+ SRAM or RLDRAM2 for specialized algorithms requiring random access performance. Ideally, the more independent banks of memory there are on the card, the better as it enables multiple market data feeds and associated processing algorithms to run concurrently. In some applications though, the lack of multiple independent memory banks can be overcome by multiplexing multiple channels onto the same memories, assuming the operating speed of these memories allows it.

Power supply: Harnessing the power of PCIe

Finding dedicated power supply lines in 1U rack mount servers can be challenging. In fact, it is a rarity in 1U servers to provide access to power supply other than through the PCI Express slots. A production-ready FPGA adapter card should be able to power itself entirely from the PCI Express power rails.

FPGA programming: Delivering long-term and hassle-free maintenance

A mechanism should exist to enable the reprogramming of the FPGA once the card is deployed in the field (for firmware updates, bug fixes), without human intervention. A typical mechanism would be to allow a new FPGA firmware image to be uploaded on the trading server remotely and programmed into the FPGA’s onboard programming Flash memory, through the PCI Express interface.

In addition to these important criteria for choosing the right FPGA card solution, there are other factors to consider:

  • Compliance and quality: For cards intended to be deployed in the field at various worldwide locations, it is important to have some level of quality assurance from the vendor in the form of compliance certifications, such as RoHS, FCC or IEC 60950 UL/CB.
  • Availability and maturity: Having cards manufactured with the latest FPGA components often means limited initial supplies and sometimes can mean that cards are mounted with early, non-production FPGA components. This could lead to  production delays and should be carefully considered.
  • Price: FPGA-based adapter cards have traditionally been viewed as an expensive route, however the benefit-to-price ratio is often well worth the expense for ultra-low latency trading and high-frequency trading applications. Additionally, as semiconductor process nodes shrink and FPGA die get smaller, prices tend to decrease with every new generation. 

 

Vendor Selection - Choosing the right partner for trading acceleration success

There are several different approaches to incorporating FPGA-based acceleration into a trading application. One key issue is understanding the different programming models for the FPGA cards and selecting the right partner for your application need.

FPGAs are intrinsically difficult to program and they typically require… FPGA design expertise! FPGA vendors, FPGA card vendors, and EDA (tool) vendors offer various methodologies and tools to simplify FPGA programming and ultimately enable software engineers to take advantage of FPGAs on their own. Unfortunately, today there is no universal solution that enables software-only developers to program and use FPGA boards, without requiring some level of FPGA expertise.

Different programming models are implemented by board vendors depending on their core competencies and business strategy:

Do it Yourself Approach: In this scenario, designing the FPGA is left entirely to the end customer. This approach requires extremely competent FPGA design skills as the customer is required to build the FPGA firmware and server-side software infrastructure, before being able to add any proprietary value-add into the FPGA.

Specialized Solutions: At the other end of the spectrum, some vendors build complex, FPGA-based trading solutions either off-the-shelf or based on customers’ requirements. Solutions designed from client specifications are required to allow some level of customization inside the FPGA, typically to enable customers to insert their own proprietary algorithms which can usually not be shared outside of the financial firm. For customers with no FPGA expertise, this can become a major hurdle for embracing FPGA technology.

 

Vendor Review - Accelize platforms for accelerated trading

Accelize is a key provider of FPGA-based programmable network platforms for the financial technology industry. Accelize platforms include advanced, FPGA network adapter cards based on the newest, production-level FPGA technology from Altera and Xilinx bundled with a comprehensive firmware and software infrastructure.  Their newest FPGA network adapters include:

  • The Accelize XP5S620LP-40G (Available now) - Based on the Altera Stratix V FPGA, it is the first low-profile FPGA computing card featuring 40Gb Ethernet and PCI Express 3.0.
  • The Accelize XP7V690LP-40G ( Coming Q3 2012) – A low-profile FPGA computing card based on the Xilinx Virtex-7 FPGA and featuring 40Gb Ethernet and PCI Express 3.0.

 

Both products are engineered to accelerate networking and financial applications and deliver the performance, configurability, and ultra-low latency necessary for high-frequency trading, smart feed handling, and real-time algorithmic trading.

Accelize enables programming of its FPGA adapter cards at different levels, based on the expertise of the user. For FPGA designers, a programmable network platform framework is available that comprises the following elements:

  • A suite of infrastructure building blocks (FPGA IP cores) to manage the card’s interfaces: PCI Express, 10G Ethernet, memory
  • A suite of server-side software modules for communicating with the card through PCI Express and accessing the card and FPGA resources: low latency PCIe device driver for Linux 32-bit and 64-bit, C API, and examples of user-mode applications

The 10G Ethernet interface building block is a key component of the platform that comprises a 10G Ethernet MAC and 10G TCP and UDP full-hardware stacks (OSI model level 4). These components enable complete TCP and UDP processing inside the FPGA.

Additionally, all of the hardware building blocks expose a standardized SoC interface bus to the user, either Avalon for Altera FPGA-based platforms or AXI4 for Xilinx FPGA-based platforms, enabling faster time-to-design as opposed to having to deal with heterogeneous interfaces.

 

Accelize - Partnering for Success

In order to address the broad spectrum of application requirements and enable each customer to focus on its areas of expertise, Accelize is working with select design partners able to provide a range of products and services that leverage Accelize FPGA network platforms to offer unprecedented value to  their mutual customers.

Impulse Accelerated Technologies (http://www.impulsec.com/), for instance, provides one of the industry’s most advanced C to FPGA compilers, enabling non-FPGA designers to take existing C code (in the financial space, typically proprietary algorithms such as trading strategies, options pricing, risk management, etc.), automatically translate the code into FPGA “RTL” code and insert that FPGA code into the infrastructure framework provided on Accelize platforms. The value for customers is obvious, especially for software only financial institutions who can therefore concentrate on their strategies without having to deal with communication and I/O interfaces inside the FPGA. 

Other partners like ENYX and xCelor are building complete trading solutions, which can be based on Accelize ultra-low latency FPGA network platforms. These solutions can be further customized based on customers’ specific requirements.

ENYX (http://www.enyxfpga.com/) for example, provides a market data acquisition solution consisting of an Accelize FPGA-based platform, which directly plugs into the trading server’s PCI-Express slot and manages the network connection to various market exchanges by receiving raw market data streams from the exchange servers and decoding the data with the best latencies.

xCelor (http://www.x-celor.com/) provides products including Market Feed Handlers (MFH) that provide parsing and order book construction within an industry-standard PCI Express based hardware solution that enables microbursts to be handled at link speed using Accelize’s FPGA-based platforms.

As FPGA technology is sparking more interest in the financial trading sector, technology and solutions providers must find ways to enable this software-centric industry to harness the power of FPGA. This requires a convergence of the hardware, firmware, and software, aimed at offering an integrated design framework, and also requires a design environment allowing non-FPGA designers to abstract hardware specificities and code the FPGA-based subsystem in a familiar, higher-level language.

Accelize ultra-low latency FPGA network platforms are a major step in that direction. The combination of leading-edge FPGA network accelerators, comprehensive infrastructure IP and server-side APIs coupled with the financial expertise and design methodologies of leading partners enables Accelize solutions to be deployed at many large banks, hedge funds, trading firms and major exchanges.

 

About the Author:

Stephane Hauradou is Vice President and co-founder of Accelize. He has worked in the fields of FPGAs and semiconductor intellectual property for over 20 years and is a recognized expert in the design and implementation of IP within an FPGA card to create value.

 

About Accelize:

Accelize is a leading provider of FPGA- based technology for the Financial Services Industry. Accelize’s ultra-low latency FPGA network platforms include a combination of leading edge FPGA processing cards along with a library of infrastructure intellectual property (IP), finance-specific IP, software APIs, and reference designs. These comprehensive and pre-validated trading frameworks provide an unprecedented time-to-market advantage by enabling financial institutions, including banks, hedge funds, specialized trading firms and exchanges, to concentrate on their value-add algorithms such as trading strategies, and proprietary IP. 

For more information, visit www.accelize.com.


Click here to download this article in PDF format


, , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , ,