About a-team Marketing Services
The knowledge platform for the financial technology industry
The knowledge platform for the financial technology industry

A-Team Insight Blogs

Q&A: Mellanox’s Asaf Wachtel on Kernel Bypass and a Future in Verbs

Subscribe to our newsletter

With a new version of its VMA software and new network cards, Mellanox is pushing down latency and jitter for both Ethernet and InfiniBand. IntelligentTradingTechnology.com spoke with the vendor’s director of business development Asaf Wachtel on the state of the art, and what’s next..

Q: For starters, can you describe briefly what your VMA software does?

A:  VMA is a dynamically linked library, which is used to offload network I/O processing from a server’s CPU.

The traffic is passed directly from the application user space to an InfiniBand or Ethernet network interface card (NIC).

It bypasses the kernel and IP stack and thus minimises context switches, buffer copies and interrupts, resulting in extremely low latency and high throughput transport performance.

Q: So what is new in the 6.0 release?

A: We have added support for our latest ConnectX-3 network adapters, which provide much higher throughput (56 gigabits on InfiniBand, 40 gigabits on Ethernet), lower latency, PCI Express Gen 3, as well as better scalability by integrating parts of the VMA functionality into the NIC’s firmware.

The combination of VMA 6.0 and ConnectX-3 delivers the lowest latency out there, but also the best scalability – maintaining the same low latency regardless of the number of cores/processes used  – and this is the real life scenario.

Q: How does 6.0 leverage your network interface cards to reduce latency?  And jitter?

A: By creating a direct path for socket communication calls to the hardware, any jitter/latency created by the operating system (which is where most jitter/latency is created) can be completely eliminated.  The new ConnectX-3 with PCIe 3.0 support provides an additional reduction in latency, as well as much faster multi-cast replication for servers running multiple processes subscribed to the same market data feed – a very common scenario.  The way the hardware I/O resources are provisioned on ConnectX-3 allows VMA to achieve the same low latency even when multiple processes are running on the same NIC.

Q: Which is lower latency and lower jitter – Ethernet or InfiniBand?

A: In general, InfiniBand provides better performance and scalability due to its architecture and capabilities.  We at Mellanox provide both options – InfiniBand and Ethernet in the same hardware, and folks can decide which to use.  Our Ethernet offering delivers the lowest possible Ethernet latency out there compared to other offerings – half versus best competitor.

Q: Do you see both 10gE and InfiniBand both existing for some years to come, or will 10gE win out?

A: Both will continue to exist.  InfiniBand share continues to increase – and we see many cases – Web 2.0, cloud, the inclusion on Windows Server 8 for SQL and more.  As we provide both options in the same interconnect solutions, we allow folks to decide what to use and when to use it – they can do time multiplexing between InfiniBand and Ethernet and match the best choice for their applications.

Q: What technology advances are you looking at for the future to reduce latency even further?  Or it that as low as it’s going to go?

A: VMA today provides the lowest latency for anyone not looking to change their application.  The Mellanox ConnectX-3 adapter can provide almost 50% lower latency when programming directly via the verbs API, and various traders as well as ISVs are taking advantage of that.  As we’ve done so far, we expect to continue and drive the latency down with our future hardware and software solutions.
 

Subscribe to our newsletter

Related content

WEBINAR

Recorded Webinar: High-Performance Networks & Low-Latency Connectivity for Trading

With financial markets becoming more complex and interconnected in today’s electronic trading environment, trading firms, exchanges, and infrastructure providers need to continually push the boundaries of network performance to stay ahead. Ultra-low latency, seamless connectivity, and resilient infrastructure are no longer just advantages – to stay competitive, they’re necessities. This webinar, part of the A-Team...

BLOG

Pico and BMLL Partner to Deliver Integrated Real-Time and Historical Market Data Solution

Pico, the global provider of financial markets technology services, has partnered with BMLL, the independent provider of historical data and analytics, to offer an integrated solution that combines real-time and historical market data to support research, liquidity analysis, and trading optimisation. The collaboration leverages Pico’s real-time data feeds and global infrastructure alongside BMLL’s Level 3,...

EVENT

Data Management Summit London

Now in its 16th year, the Data Management Summit (DMS) in London brings together the European capital markets enterprise data management community, to explore how data strategy is evolving to drive business outcomes and speed to market in changing times.

GUIDE

AI in Capital Markets: Practical Insight for a Transforming Industry – Free Handbook

AI is no longer on the horizon – it’s embedded in the infrastructure of modern capital markets. But separating real impact from inflated promises requires a grounded, practical understanding. The AI in Capital Markets Handbook 2025 provides exactly that. Designed for data-driven professionals across the trade life-cycle, compliance, infrastructure, and strategy, this handbook goes beyond...