The quest to build the FastNIC

Martyn Warwick
By Martyn Warwick

Oct 7, 2019

via Flickr © ndrwfgg (CC BY 2.0)

via Flickr © ndrwfgg (CC BY 2.0)

  • Old Network Interface Cards can't keep pace with today's data processors
  • Current network stack protocols are no longer up to the job
  • Data bottlenecks slowing down vitally important supercomputers and data centres
  • Rebuilding the stack to remove intrinsic structural data transport speed limits

The search is on to find a solution to alleviate the data choke-points and bottlenecks that are bedevilling supercomputing and data centres. Long-established network stack protocols and interfaces are old-fashioned and slow. They cannot keep pace with today’s high-end processors and parallel processing capabilities. Something new is needed.

Fifty years (and a wee bit more) ago the Advanced Research Projects Agency Network (ARPANET), funded by the US Department of Defense became operational. It was a pioneer packet-switching network that allowed multiple computers to communicate across a single network and was the very first one to implement TCP/IP (Transmission Control Protocol and Internet Protocol protocol) and that allowed the construction of a network of networks that eventually became the Internet.

Half a century on, ARPA, under its incarnation as DARPA, (the Defense Advanced Research Projects Agency), a department of the Pentagon, is endeavouring to modernise and reconstruct both the network stack and interfaces that can no longer keep pace with today's high-end processors.

We all know how frustrating it is when a network connection degrades from a sprint to a crawl and work slows to less than a snail's pace. And that' s just with a smartphone or a laptop. Well, that is also happening right now with supercomputers.

Vast arrays of interlinked computer cores and thousands upon thousands of servers and GPUs are processing at incredibly high speeds but are then more or less stopped in their tracks as they wait for old and overwhelmed network interfaces to stay in sync and trundle slowly through the ever-increasing amounts data traffic they have to deal with. This confluence between Network Interface Cards (NICs) and other elements of a network is an intrinsic structural barrier that limits the speed with which data can be shared between different computing resources.

Hence the search for new NIC technologies with network equipment vendors being asked to come up with fresh ideas and designs that would result in a hundred-fold increase in interface speeds and thus greatly boost network performance and clear the data traffic bottlenecks that are clogging networks. The new initiative, "FastNIC", will (hopefully) result in the complete reinvention and rebuilding of the network stack from the application level to the system software layers running on top of much faster hardware.

As Dr. Jonathan Smith of DARPA says, “The true bottleneck for processor throughput is the network interface used to connect a machine to an external network, such as an Ethernet, therefore severely limiting a processor’s data ingest capability."

By the way, the origins of Ethernet lie 56 years in the past and the technology fully developed in the 1970s, has been in commercial use since1980. In terms of the many generations of computing systems and technologies that have come and gone over the past 39 years, Ethernet is not so much venerable as the Methuselah of computer networking.

New FastNIC software libraries will be open source

As Dr. Smith says, “There is a lot of expense and complexity involved in building a network stack  It starts with the hardware; if you cannot get that right, you are stuck. Software can’t make things faster than the physical layer will allow so we have to first change the physical layer.”

The DARPA scientists reckon that computer processors and memories generally function at a speed of some 10^14 bits per second, as do networking hardware components such as switches and fibre optic cable. That's well into terabit territory. However, current NICs operate in the gigabit range and therein lies the problem. It's like trying to pump the contents of a massive reservoir down a garden hose. It would be so slow as to be useless.

That's why DARPA is challenging research scientists to design and demonstrate 10 Tbps NICs using extant or already planned-for hardware interfaces. The main requirement is that the cards must marry with servers via at least one (and preferably more) industry standard interface points. Once that is achieved the next step will be the development of a software system capable of managing the hardware -  and that will not be easy. It will however be expensive and time-consuming.

No doubt some of the big vendors would wish to head off down the path of proprietary systems with the intent to lock-down the technologies and lock-in customers. After all, historically that's what they have always tried to do - but the good news is that the software libraries that will be created by FastNIC will all be open source so the emergent standard will not be the exclusive proprietary intellectual property of vendors or even the mighty US Defense Department.

No timetable for the development of FastNICs has been proposed and it is not envisaged that one will be imposed. However, rumour has it that DARPA is hoping that a viable solution might be available within three years. If that happens things are going to change, Fast, by 2023.

Email Newsletters

Sign up to receive TelecomTV's top news and videos, plus exclusive subscriber-only content direct to your inbox.