No announcement yet.

Linux' TCP/IP is obsolete

  • Filter
  • Time
  • Show
Clear All
new posts

  • Linux' TCP/IP is obsolete

    1. In-Kernel TCP/IP vs. Userspace TCP/IP
    I had an epiphany one day when I realized that the kernel is nothing but a library with an expensive calling convention.
    The kernel has always been the problem. User-space is where it's at.

    Todays high-performance network applications spend a significant portion of CPU cycles inside the Kernel for TCP processing, for lighttpd you can assume, these are around 80%. With TCP/IP being a bottleneck, optimizing the Application atop it make little sense. Another bunch of cycles is wasted for system call overhead. As of today, there's a movement to pull the TCP processing into the user mode. A actually usable prototype has already been created.

    An Open efford of the industry is being made to create an fast TCP/IP implementation in the usermode.

    2. Faithful Implementation of standards vs. Slapdash work

    Well, the linux guys have always been great in inventing their own standards, but following other standards? Linux implements the POSIX standard with only little faith, why should it be different with the Internet protocol suite.

    3. Desktop Model vs. Server Networking

    On a Desktop, we have multiple applications, such as, for example: Steam syncing with one users Steam account, a browser pulling websites from webservers, a Bittorrent client, a Gnutella client, etc. They need proper isolation from each other. On the other hand, an internet server is usually dedicated to a single task (for example, Webserver or Database server), and it is speed what matters here.

  • #2
    Nokia, ARM, Enea craft new TCP/IP stack for the cloud

    Open source 'OpenFastPath' unloads comms from the kernel

    A group of major vendors has put forward an open source TCP/IP stack they say is designed to reinvigorate the ancient and rather crusty protocol.

    Nokia, ARM, and Enea are offering up both code and tutorials here for their OpenFastPath user-space TCP/IP implementation.

    As The Register has previously noted, user-space networking is designed to get TCP/IP out of the kernel space, for two reasons: kernels have absorbed a lot of code over the years; and using the kernel for packet processing involves extra operations to get packets into memory, pass them to the kernel, and push them back out to the interface.

    The BBC, for example, has crafted a user-space stack so its video farms can push out the 340,000 packets per second needed for high-definition video, and a Swiss company called Teclo is using a user-space stack to improve mobile network processing.

    This will be a huge deal for the next generation of Server applications. If you are a developer of server Applications, please be aware of OFP and similar projects. The time of garage-work servers is over! The future needs a software-architekture for servers, that goes beyond the desktop-model.


    • #3
      Many believe that a critical success factor for 5G will be a fully revamped TCP/IP stack, optimized for the massively varied use cases of the next mobile generation, for cloud services, and for virtualization and software-defined networking (SDN). This is the goal of the new OpenFastPath (OFP) Foundation, founded by Nokia Networks, ARM and industrial IT services player Enea. This aims to create an open source TCP/IP stack which can accelerate the move towards SDN in carrier and enterprise networks. Other sign-ups include AMD, Cavium, Freescale, Hewlett Packard Enterprise and the ARM-associated open source initiative, Linaro.

      Nokia and its allies hope to accelerate the platforms to create open but secure network applications, which harness IP packet processing and support a whole range of use cases, some requiring very high throughput, others ultra-low latency (or both).

      It looks like future servers will run on OFP.