Effective systems are only partly about hardware. Focusing on packet processing, 6WIND provides software that accelerates network operations. High Frequency Traders spoke to 6WIND’s CEO, Eric Carmès, and VP of Marketing, Charlie Ashton, to find out more about what the company’s solutions can do and what part they are playing in this rapidly evolving space.
HFT: First of all, can you give us a little background on the company?
Charlie: 6WIND has been in business since 2000. It was originally founded by a number of networking experts who had previously been at Thales. The initial focus of the company was on solutions around the latest Internet Protocol version (IPv6) networking, hence the company name. By 2005-2006 the team saw a big opportunity in the area of high performance networking software for the multi-core processors that were just coming on to the market and starting to be used in high-end networking systems.
By taking full advantage of the multiple processing elements found in a multi-core CPU and being smart about how individual packets are handled, we are able to significantly improve the packet processing throughput of a network system.
That meant the focus of the company really switched about six years ago to where we are now. Our whole focus is on packet processing solutions that are optimised for multi-core platforms. By taking full advantage of the multiple processing elements found in a multi-core CPU and being smart about how individual packets are handled, we are able to significantly improve the packet processing throughput of a network system.
We have very close relationships with most of the leading multi-core processor suppliers, including Cavium, Freescale, Intel, NetLogic Microsystems and Tilera. We’ve been very successful in the mobile infrastructure market where about two-thirds of the Tier 1 manufacturers use our software in their core networking equipment.
Over the past couple of years we’ve expanded into other performance-oriented networking applications such as Enterprise Networking, Network Security and Deep Packet Inspection. High frequency trading has become another focus area for us. We found that our software was very well suited to meet the low-latency requirements of networking systems supporting computerized trading applications.
HFT: What technical challenges result from the growing trend towards the use of multi-core platforms in financial applications?
Charlie: There are a couple of major challenges when you start to do high performance networking on multi-core processors. The first challenge is that standard operating systems such as Linux are extremely inefficient for doing packet processing. That’s because every single packet has to be processed through the multiple Internet Protocol layers, called ‘the stack’, and therefore is subjected to all kinds of latencies and processing-induced overhead leading to a major performance hit. One of the things that we address through the architecture of our software is we dedicate certain cores on the multi-core processor to only do packet processing and essentially to run outside of the operating system. With this approach we’re able to get at least an order of magnitude performance improvement for packet processing and reduce protocol processing latencies, which is extremely important to HFT applications.
There are many other challenges associated with supporting modern multi-core technology. A particularly tricky example is in providing support for the specialized on-chip accelerators that are included in each vendor’s products. Taking full advantage of these accelerators provides another level of performance improvement important to HFT applications, but requires that we must support the wide variety of accelerators used across all of these architectures.
Eric: The vision of the company is really all about software-based networking. When we discovered the multi-core architecture five or six years ago, I think we understood earlier than the competition that it would be possible to achieve very high performance with software architecture designed for multi-core processors. Of course packet processing has to be specifically designed to extract the highest possible performance. We have demonstrated this in mobile infrastructure where the stress on the performance and capacity is very significant in the new generation of mobile networks. It’s the same concept with HFT. It’s the idea that an efficient software solution can really address the requirements in different markets. Of course, HFT has requirements different from mobile infrastructure. It’s more a question of latency versus managing a very high bandwidth, but it’s the same idea. We are able to provide a software solution that offers not only performance but also flexibility that satisfies the stringent network requirements. That’s why we have good credibility in the HFT market.
Being able to develop software outside of the operating system makes it possible to minimise the latency around packet processing and fine-tune the distribution of the incoming packets to cores. It means that you can be sure that an incoming packet can be allocated to an available core. By doing this outside the OS and managing the allocation of packets to cores we have a very useful solution to minimise latency.
HFT: One of the 6WINDGate solutions features is limited interaction with the operating system. What advantages does this bring?
Eric: We provide two levels of improvement. Generally multi-core processors are delivered with what they call an ‘executive’ which is in fact very low-level software element that provides limited features compared to an operating system, but very optimised in terms of packet network drivers, allocation of packets to the cores, timers and memory management. It’s very useful because compared to an OS, you don’t have these problems regarding latency. Being able to develop software outside of the operating system makes it possible to minimise the latency around packet processing and fine-tune the distribution of the incoming packets to cores. It means that you can be sure that an incoming packet can be allocated to an available core. By doing this outside the OS and managing the allocation of packets to cores we have a very useful solution to minimise latency.
HFT: Do you have any examples where implementation of your technology has brought benefits to trading organisations?
Eric: Without disclosing any specific names we have a number of deployments in financial organisations. Our solution is used to implement TCP proxy de-capsulation and encapsulation of the protocol. One customer, an HFT platform provider, gave us a latency budget for this networking layer, which we then implemented. One very interesting feature is that our technology is completely implemented on an inexpensive Cavium Octeon processor running at 750 Mhz. With this we were able to implement the TCP proxy running at less than 10 microseconds. The new generation of Cavium processors (Octeon II) is running much faster (up to 1.5Ghz), which means that when necessary we will be able to significantly improve the performance of our TCP proxy function thanks to the performance of the processor technology.
Coming from the mobile infrastructure market, we have already implemented a number of protocols such as firewall, security (IPsec) and so on. These security features are going to be requirements for HFT in the future, so thanks to our background in mobile infrastructure, we are well placed to service these requirements.
HFT: What advantages does the use of open API bring to your solutions?
Eric: Latency for HFT is a system problem, which means that you have the networking part, but also the application has to be optimised to minimise latency at the application level. So when we provide a networking layer, it is very important for our customers to have a standard API to be able to integrate the application and network as quickly as possible. It also helps to ensure that the integration of the application is optimal, hence eliminating latency. So we have a very straightforward API that makes it possible to integrate easily with financial applications. For example, the customer I mentioned earlier was able to implement the application within a couple of weeks. We also have some development tools that make it possible to develop the application at the user level and then to recompile the solution on the target. The open API helps make the whole process more efficient.
HFT: The low-latency space is moving very quickly. How is it possible to ensure that solutions are future-proof?
Eric: Our technology has a number of benefits that can be used to ensure we are always up to date. First of all, we have a portable solution which runs on all the industry leading multi-core platforms. As those platforms become quicker and more powerful we will automatically benefit. There is also a trend in the industry for increasing numbers of cores. When we have more cores we will be able to distribute the packets to a larger number of cores, which means that within a latency budget we will be able to manage a larger bandwidth. This means that when the number of transactions increases we will be able to allocate them to more cores. Our technology can be implemented on a single processor board but also be extended over several boards if you need more cores to process the packets. Regarding networking features, we will be able to use more processing capability to implement more sophisticated features while also keeping the latency budget at the same level. Thanks to technology improvements we will have more processor cycles and we will be able to do more. We are very well positioned to benefit from processor technology improvement and fulfil future HFT requirements.