Re: TCP customization for httpd

From: Joe Touch ([email protected])
Date: Thu Feb 01 2001 - 17:14:04 EST

  • Next message: Andrew-Mark Pether: "Re: TCP customization for httpd"

    Thomas Bohn wrote:
    >
    > All:
    >
    > Please excuse this post if inappropriate.
    >
    > Am seeking any interest in an opportunity to test research
    > projects/products in a working satellite environment: am working on a US
    > Federal telehealth project in the state of Alaska -- we have a statewide
    > network interconnected by a combination of satellite-based and terrestrial
    > T1 and sub-DS1 links. We have developed a web-based application for
    > digital acquisition of medical imagery in remote areas, with a
    > store-and-forward function for referral of medical data to distant
    > providers/specialists. The application has been developed on Win2K
    > platforms (both client and server) using ColdFusion middleware and SQL
    > database backends. I am currently porting the application to Linux.
    >
    > We have run into the well-known snag with sat links and TCP: short bursty
    > data transfers over TCP don't do so well on high-band/long delay nets.
    >
    > I am seeking information related to Open Source TCP stack modifications
    > for a web proxy to be placed on the web server LAN in order to accelerate
    > the transfer of data out over the WAN to distant client machines. We do
    > *NOT* want to modify the stack on the webserver itself, but on an
    > accelerating proxy for outlying satnet clients.
    >
    > Our network is private, e.g. is NOT connected to the public Internet. We
    > are built out on RFC 1918 address space, use Cisco network elements and
    > standard transport offerings from Alaska-based space segment providers.
    >
    > If anyone on this list has an interest in testing within such an
    > environment (whether your solution is commercial or non-commercial) we
    > would like to speak with you. We are, as stated, seeking non-commercial
    > Open Source variants of standard TCP implementations -- primarily on the
    > sender side. Link layer framing is typically standard serial, as
    > ATM-based cell protocol offerings in the Alaska market were recently
    > pulled by AT&T (e.g. we use HDLC, PPP, etc.)

    Sounds like what you want to do is:

            - run HTTP 1.1 (or at least +persistent connections, which predates 1.1)

            - turn TCP's congestion management
                    one way to do this using private space is to have the endpoints
                    on the same subnet. that causes many implementations to turn
                    off congestion management

            - use large MTUs (the same trick - same subnet will trigger use of link MTU
              rather than path MTU discovery, but be careful of black holes if that size
              is too large -- it can be reset at the interfaces themselves)

            - make sure your socket buffers, etc are tuned properly (see www.psc.edu)

    Note that none of these necessitate the use of a proxy.
    Some can be implemented at the proxy, but that runs the risk
    of having the proxy be the bottleneck.

    Joe



    This archive was generated by hypermail 2b29 : Thu Feb 01 2001 - 18:02:40 EST