Part 1 of a 3-part Series
Low latency has long been an important consideration in many market segments of computing. This includes the Financial Services Industry where a 1-usec advantage has been said to be worth $100 million a year. In the Broadcast, Media and Entertainment segment, increasing video quality video including the latest Ultra High Definition formats require high bandwidth, low jitter (variance in packet delays) networks – so latency must be kept as small as possible.
Applications like these have brought low latency concerns to the forefront as an industry challenge and I’ll be addressing them in this three part blog series. Aristotle said, “The whole is greater than the sum of its parts”, however, when it comes to Latency, there is a variety of approaches that can applied with each helping to reduce overall latency. In this three part series blog we’ll look at several of the technologies Mellanox deploys to offer high performance, low latency Ethernet solutions. In the first segment, I’ll discuss what Mellanox is doing to stay at the forefront of low latency solutions and how we are resolving challenges for the Financial Services Industry. I’ll expand on this topic in the following two segments and discuss the adapter and switch’s role in keeping latencies low. Finishing the series will be discussion around messaging acceleration software and offloads. In the case of Mellanox, we offer several approaches to lowering latencies which can make the overall latency less than the sum of all parts.
Research and Development
Sustaining low latencies takes focus and continued investment in research and development. One of the ways Mellanox places focus on in-house research and development efforts is by hosting and conducting multiple regional internal Hackathons each year. At these events, software development teams are tasked to collaborate and come up with solutions that solve specific industry challenges. This year’s focus was to lower latencies and increase message rates. This was our 8th annual hackathon and we did just that – we set a new standard in reaching low latencies. The starting goal was to break previous latency barriers with new developments that can be implemented in future product and software releases, with the expectation that these solutions will benefit the Financial Services Industry and other markets that rely on low latency solutions. Designing a low latency infrastructure is complex undertaking, especially when the goal is to avoid trading some degree of efficiency for improved latency gains. The teams were up for the challenge and had one advantage, designed around our newest Mellanox ConnectX®-5 Network Adapters which are capable of up to 100 Gb/s Ethernet speeds and 200 million messages per second.
The results of this year’s Hackathon were greater than what could have ever been expected. Latencies were reduced against the latest competitive devices at all data points. Furthermore, by thinking outside the box, the winning team demonstrated a whole new way of handling traffic and managed to cut latency down to the hundreds of nanoseconds. By comparison, current products on the market today measure latency in microseconds.
In the Hackathon, our teams demonstrated two drastic latency improvements over competing offerings. One team made use of on-chip buffering, a newly introduced technology available on the ConnectX-5 network adaptors, to cut latency by as much as 33 percent. This improvement will be rolled into our upcoming software release in June of 2017. Second, by using some creative concepts, another team was able to reduce latencies to around 600 nSec. This superior performance demonstrated our ability to lower latency by as much as 70 percent compared to all other NICs on the market. The idea for this breakthrough was first leveraged by Mellanox for use within the High Performance Computing (HPC) market and our InfiniBand products. The concept is based on extending and advancing the in-network computing capabilities of the Mellanox NICs, a new feature within our ConnectX-5 adapters. And while this feature is not yet available to the general market, it will be available later this year.
By placing the focus of the Hackathon on achieving greater performance through lowering latencies, the Mellanox teams were able to design specially tailored performance enhancements that directly impact latency and ultimately improve performance for applications that mandate ultra-low latency. Through a creative and innovative approach, each team was able to reduce latencies and demonstrate performance gains against competitive solutions. With these improvements, organizations have an opportunity to design a network that can easily exceed performance and latency requirements – better than anything on the market today. These enhancement are especially important for time-critical decision making application such as the financial market segment which can use these to their advantage in Exchanges, Trading Platforms and High Frequency Trading (HFT) environments. Ultimately, these factors translate into higher profits for these financial organizations.
Continued Latency Reductions Throughout 2017
As the Financial Service Industry races to the finish line, intent on completing analytical algorithms faster than their competitors, the innovative ideas from the Mellanox Hackathon will continue to emerge throughout 2017 in future software releases and will help to reach new levels of latencies reductions for these firms. This is sure to drive Mellanox even further as a leader in the high-performance market that continues to break existing latency barriers.
In the next blog segment from this three part series, I’ll discuss how Mellanox is further addressing latency and message rates with advancements in switching and adapters specifically for the Financial Services Industry. Following that segment, I’ll discuss the Mellanox software acceleration solution and how this is being used in other market segments too like Media and Entertainment where technology struggles to keep pace with increasing video resolution. I encourage you to check back and read the final two parts of this three part series.