[lkml]   [2000]   [Feb]   [7]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
    SubjectRe: Gigabit Linux Server Bottlenecks

    You have pointed out, very delicately, a major limitation of Ethernet as
    it is implemented today. I find myself constantly arguing against
    implementation of gig Ethernet due to the high CPU cost on the host
    system. Trying to fill a gig pipe with 1,514 byte packets is, IMHO,
    maddening to the CPU. As I recall, the host must be able to process
    around 82,000 packets (read "CPU interrupts") per second to fill the
    pipe unidirectionally. Try this in a full duplex setting with a 2 gig
    pipe to fill and you (at least) double the amount of work the CPU needs
    to accomplish every second.

    Since no one seems to want to jump on the 8K jumbo frames bandwagon
    Alteon has embraced, current technology and CPU speeds seem to indicate
    more "bang for the buck" with a couple of trunked Fast Ethernet links.
    The last thing one wants to do, no matter what NOS you choose, is
    overrun the host CPU. Depending on the environment, it may even be less
    expensive when all things are considered to implement a Token Ring

    Personally, I recommend one of the following to my clients.

    1. If Token Ring / Fast Token Ring is implemented anywhere in the
    enterprise, keep it. If the client absolutely wants client segments as
    Ethernet, move the Token Ring hardware to the data center and implement
    switched 10 (or 100 iff Fast TR in the DC) Ethernet to the clients.

    2. If there are no implementations of Token Ring, go gig with Jumbo
    Frames in the DC and spend the $$$ on the edge devices to convert the
    packets back to 1514 bytes. Client segments still use "Classic" Ethernet
    packets and don't need the high end ACEnic's.

    Either of these scenarios provides an adequate solution to getting the
    data out of the DC as quickly as possible.

    Craig Rich wrote:
    > Hello:
    > I hope no one minds the cross posting.
    > I'm a new subscriber to these lists, trying to do some research
    > on
    > Gigabit Ethernet. My company develops Etherent silicon (MACs mainly),
    > and we are in the process of developing a Gigabit Ethernet MAC. We are
    > trying to get an understanding of what features we could add to our next
    > Gigabit Ethernet product in order to maximize performance. I have some
    > experience with Linux (I've written the Linux device drivers for our
    > 10/100 and Gigabit Ethernet MACs), but my overall Linux experience is
    > not great (only been working with Linux for about 1 year now).
    > We feel that Gigabit Ethernet can put a large load on a host
    > system
    > (i.e. server) processor just from processing Ethernet frames, and TCP/IP
    > datagrams. We have some ideas on what we could do to offload the host
    > processor and improve performance, and I was hoping list subscribers
    > could help me with the following questions:
    > 0) I'm unsure if this list is the best place for feedback. If it is
    > not, any suggestions on other lists or information sources?
    > 1) Is there much experience out there with Gigabit Ethernet
    > in Linux servers? If so, which Gigabit Ethernet products (NICs,
    > switches, routers) are utilitzed?
    > 2) Where does the performance (IP throughput for example) peak for
    > a Gigabit enabled Linux server? What is the average throughput?
    > 3) What are the main bottlenecks with regard to network performance
    > (IP throughput) for Gigabit enabled servers? By bottlenecks I
    > mean what datagram or fram processing requirements take up the
    > most CPU time when processing streams of Gigabit speed data?
    > 4) How popular are IPSec, and/or other network security protocols (SSL
    > for example) in Gigabit Linux servers?
    > 5) If a NIC was available which could provide hardware functions such
    > as TCP segmentation, encryption/de-cryption, IP datagram signature
    > calculation (i.e. MD5, SHA-1), and/or other datagram processing
    > functions, are there any distributions of Linux today which could
    > take advantage of these off load functions or would the Linux
    > networking code have to be largely re-written?
    > Thanks again for reading through this post. Any input or
    > pointers to
    > information would be greatly appreciated.
    > --
    > Craig Rich
    > Senior Systems Engineer
    > Sundance Technology Inc.
    > 1485 Saratoga Avenue tel: 408 873 4117 x107
    > Suite 200 fax: 408 996 7064
    > San Jose, CA, USA 95129
    > -
    > To unsubscribe from this list: send the line "unsubscribe linux-net" in
    > the body of a message to

    John LeMay Jr.
    Senior Enterprise Consultant
    NJMC, LLC.

    The rules have changed... Get paid to surf the web!!!
    n:LeMay Jr.;John
    tel;fax:732 244-3351
    tel;work:732 785-2525
    org:NJMC, LLC.
    adr:;;10 Fella St.;Howell;NJ;07731;USA
    title:Senior Enterprise Consultant
    fn:John LeMay Jr.
     \ /
      Last update: 2005-03-22 13:56    [W:0.029 / U:7.088 seconds]
    ©2003-2016 Jasper Spaans. hosted at Digital OceanAdvertise on this site