[ad_1]
Having a slow connection is always frustrating, but imagine what supercomputers are feeling. All of these cores perform all kinds of light-fast processing, but they all expect an outdated network interface to stay in sync. DARPA do not like it. So, DARPA wants to change it – especially by making a New network interface a hundred times faster.
Here is the problem. According to estimates by DARPA, the processors and memory of a computer or server can operate at a speed of about 10 ^ 14 bits per second – which is quite in the terabits region – and the Network hardware, like switches and fiber, is capable of the same.
"The real bottleneck for processor throughput is the network interface used to connect a machine to an external network, such as Ethernet, thus severely limiting the power supply of a processor's data, "said Jonathan Smith of DARPA in an article published by the agency on the project. (Underlined by me.)
This network interface is generally in the form of a card (transformed into a network card), accepts the data from the network and transmits it to the computer systems, or vice versa. Unfortunately, its performance is generally more in the gigabit range.
This delta between the network adapter and other network components imposes a fundamental limitation on the speed of information sharing between different computing units, such as hundreds or thousands of servers and GPUs that make up supercomputers and data centers. The sooner a unit can share its information with another, the faster it can move on to the next task.
Think of it like this: you operate an apple farm, and every apple needs to be inspected and polished. You have people inspecting apples and people polishing them, and both can make 14 apples a minute. But the conveyor belts between departments only carry 10 apples per minute. You can see how things are going to accumulate and how frustrating it would be for everyone involved!
With the FastNIC program, DARPA wants to "reinvent the network stack" and improve throughput 100 times. After all, if they manage to solve this problem, their supercomputers will have a huge advantage over others in the world, especially in China, which rivals the US in the computer arena. high performance for years. But it will not be easy.
"Building a network stack requires a lot of expense and complexity," said Smith, whose first project will be to physically redefine the interface. "It starts with the material; if you can not get that right, you're stuck. The software can not do things faster than the physical layer allows. So we have to change the physical layer. "
The other main part will be, of course, to redo the software side to cope with the huge increase in the scale of the data that the interface will have to manage. Even a 2x or 4x change would require systematic improvements; 100x will roughly involve a complete overhaul of the system.
The researchers at the agency – supported, of course, by all the private sector actors who want to participate, so to speak – aim to demonstrate a connection of 10 terabits, although there is still no timetable . But the good news at the moment is that all software libraries created by FastNIC will be open-source. This standard will not be limited to the proprietary systems of the Department of Defense.
FastNIC is just getting started, so forget it for the moment and we'll let you know when DARPA decodes the code in a year or three.
[ad_2]
Source link