The leading knowledge platform for the financial technology industry
The leading knowledge platform for the financial technology industry

A-Team Insight Blogs

Messaging Platforms and Hardware Choices: General Purpose or Accelerated?

Hardware-only vs hybrid vs software-only approaches to handling high data volumes with ultra low latency – Is there a clear winning approach or is the solution dependent on the specific application?

Everybody uses hardware for messaging. The real question is: general-purpose or special-purpose hardware? Even special-purpose hardware starts life as software (e.g. VHDL code) which is compiled to design FPGAs. The software for such machines is generally written and maintained by the vendor of the machines rather than by their owners.

General-purpose hardware typically means an x86 processor running a common operating system with a very rich development environment. Although vendors may have written the operating system – and perhaps the messaging libraries – the key application code the machine executes is generally written by its owner.

Usually, special-purpose hardware is faster than general-purpose hardware, but is more expensive and more difficult to change due to rudimentary development tools. 29West, who licenses messaging libraries that run on general purpose hardware, has watched developments in the world of special-purpose hardware with interest since it potentially could help some of the applications where our products are widely deployed.

One place where special-purpose hardware might make sense is in systems where it can be used for all components of the latency-sensitive messaging path. Imagine if algorithmic trading strategies were implemented in special-purpose hardware machines with no need to use general-purpose hardware at all.

Compared with strategies implemented on general-purpose hardware, such machines would cost more and take longer to develop, but there’s no question that they would have lower latency between the arrival of market data and the placement of an order. We’re not aware of anyone who has built such machines, though, since it appears to be a formidable undertaking.

It’s also clear that performance improves when you replace a software message router process running on general-purpose hardware with a hardware message router (made from special-purpose hardware). But many companies we talk with ask, why not take this idea one step further and use existing general-purpose network routers and switches to do message routing?

Most any infrastructure-grade router or switch built in the last five years has the ability to copy multicast messages at wire-speed. Similarly, they can filter out unwanted messages on different multicast groups. It’s important to note that a growing number of customers are taking advantage of software that runs on inexpensive, general purpose CPU hardware and allows applications to tap the wire speed performance of existing routers and switches.

Are there currently deployments of special hardware in systems driven by market data? We do recognise one task for which special-purpose hardware may be well-suited: that’s the TCP fan out and filtering case. The job of turning a single TCP input stream into several TCP output streams cannot be done by the hardware in today’s switches and routers.

Further, they are incapable of tracking the interests of individual receivers and selectively filtering out uninteresting messages. Software messaging systems can perform these tasks using general-purpose hardware. Special-purpose hardware will be faster, but generally more costly and difficult to change.

We believe that general-purpose CPUs, existing network hardware, and messaging software make the best combination for the vast majority of financial market applications we’ve seen. The only exceptions we know are on networks that cannot support multicast addressing or those where the rich development environments of general-purpose hardware are not required for any component of the system.

What can be proven? It’s easy to see that any message starting in software in one machine and ending in software in another machine while passing though a set of network switches and routers will arrive sooner than one that goes through all of the above plus a hardware message router.

A hardware message router can only add latency when compared to the direct path. Software messaging systems are always faster than a hardware message router in this regard.

Related content


Upcoming Webinar: Integrating Intelligent Machine Readable News

Date: 30 November 2021 Time: 10:00am ET / 3:00pm London / 4:00pm CET Duration: 50 minutes Intelligent machine readable news is a powerful tool in the arsenals of trading and investment firms seeking competitive advantage. It turns unstructured data into actionable insight and can be used, for example, to uncover market trends, identify correlations and...


IPSX Creates Integrated Technology Framework Ahead of First Admission

UK-based IPSX, which describes itself as the world’s first regulated stock exchange dedicated to real estate assets, has pulled together an integrated technology environment and a set of standardised data workflows ahead of its first ‘Intention to Float’ (ITF) on its IPSX Wholesale market, scheduled for next month. IPSX has been working with a plethora...


RegTech Summit New York City

Now in its 5th year, the RegTech Summit in NYC explores how the North American financial services industry can leverage technology to drive innovation, cut costs and support regulatory change.


Trading Regulations Handbook 2021

In these unprecedented times, a carefully crafted trading infrastructure is crucial for capital markets participants. Yet, the impact of trading regulations on infrastructure can be difficult to manage. The Trading Regulations Handbook 2021 can help. It provides all the essentials you need to know about regulations impacting trading operations, data and technology. A-Team Group’s Trading...