DARPA strives to make
networks 100 times faster thanks to FastNIC
A slow connection is always frustrating,
but imagine how supercomputers feel. All of these cores perform all kinds of
processing at lightning speed, but in the end everyone is waiting for the
outdated network interface to keep synchronization. DARPA doesn't like it.
That's why DARPA wants to change it - in particular by creating a new network
interface a hundred times faster.
The problem is this. As DARPA estimates,
processors and memory on a computer or server can generally run at about 10 ^
14 bits per second - that's convenient in a terabit region - and network
equipment such as switches and fiber optics is able to do the same.
"The real bottleneck in processor
bandwidth is the network interface used to connect the machine to an external
network, such as Ethernet, which significantly limits the processor's data
processing capabilities," explained Jonathan Smith of DARPA in a press
release about the project. (Highlight mine.)
This network interface usually takes the
form of a card (which makes it a network card) and supports receiving data from
the network and transferring it to your own computer systems or vice versa.
Unfortunately, its performance is usually higher in the gigabyte range.
This difference between the network adapter
and other network components means a basic limit on the speed of sharing
information between different computing units - such as hundreds or thousands
of servers and GPUs that make up supercomputers and data centers. The faster
one unit can share information with another, the faster it can go to the next
task.
Think about it this way: you run an apple
farm, and each apple must be checked and refined. People check apples and
polish apples, and they can both make 14 apples a minute. But conveyors between
departments carry only 10 apples per minute. You can see how everything will
turn out and how frustrating it would be for everyone involved!
Thanks to the FastNIC program, DARPA wants
to "reinvent the network stack" and increase throughput 100 times.
After all, if they manage to solve this problem, their supercomputers will have
a huge advantage over others in the world, in particular those in China that
have been competing with the US for years in the arena of high-performance
computers. But it won't be easy.
The second main part will obviously be the
modification of the software site to cope with the huge increase in data scale
that the interface will have to handle. Even a 2x or 4x change would require
systematic improvements; 100x will require a thorough overhaul of the system.
Researchers from the agency - supported, of
course, by private individuals from the industry who want to unite, so to speak
- strive to demonstrate a connection of 10 terabits, although there is no time
line yet. The good news is, however, that all software libraries created by
FastNIC will be open source, so this standard will not be limited to the
Department of Defense's proprietary systems. FastNIC is just starting, so for
now, forget it and we'll let you know when DARPA breaks the code within a year
or three.
No comments:
Post a Comment