HTTP Connection Management (draft-ietf-http-connection-00.txt) (fwd)

From: Miguel A.L. Paraz <map@dont-contact.us>
Date: Wed, 26 Mar 1997 20:41:44 +0800 (HKT)

Hi,

Picked this up from the Apache list. Applicable?

---------- Forwarded message ----------
Date: Wed, 26 Mar 97 00:18:08 -0500
From: jg@zorch.w3.org
To: internet-drafts@ietf.org
Cc: masinter@pa.dec.com, rstevens@kohala.com, freier@netscape.com,
    http-wg@cuckoo.hpl.hp.com, jg@pa.dec.com
Subject: HTTP Connection Management (draft-ietf-http-connection-00.txt)
Resent-Date: Wed, 26 Mar 1997 00:37:49 -0500 (EST)
Resent-From: http-wg@cuckoo.hpl.hp.com

Alan Freier and I (and others) have been worried about the fact that RFC2068
is silent on much of the rational and implementation detail required for
successful implementation of connection management policy for HTTP persistent
connections in HTTP/1.1. Alan goaded me into drafting this with him
a few weeks ago, and we now have something presentable for working group
comments. I'm submitting this to the ID drafts editor.

I'll make this available in HTML sometime next week, but you'll have
to live with plain-text for the moment.
                                - Jim Gettys
=========

HTTP Working Group J. Gettys, Digital Equipment Corporation
INTERNET-DRAFT A. Freier, Netscape Communications Corporation
Expires September 26, 1997 March 26, 1997

                       HTTP Connection Management

                   draft-ietf-http-connection-00.txt

Status of This Memo

   This document is an Internet-Draft. Internet-Drafts are working
   documents of the Internet Engineering Task Force (IETF), its areas,
   and its working groups. Note that other groups may also distribute
   working documents as Internet-Drafts.

   Internet-Drafts are draft documents valid for a maximum of six months
   and may be updated, replaced, or obsoleted by other documents at any
   time. It is inappropriate to use Internet-Drafts as reference
   material or to cite them other than as "work in progress."

   To learn the current status of any Internet-Draft, please check the
   "1id-abstracts.txt" listing contained in the Internet-Drafts Shadow
   Directories on ftp.is.co.za (Africa), nic.nordu.net (Europe),
   munnari.oz.au (Pacific Rim), ds.internic.net (US East Coast), or
   ftp.isi.edu (US West Coast).

   Distribution of this document is unlimited. Please send comments to
   the HTTP working group at "http-wg@cuckoo.hpl.hp.com". Discussions
   of the working group are archived at
   "http://www.ics.uci.edu/pub/ietf/http/". General discussions about
   HTTP and the applications which use HTTP should take place on the
   "www-talk@w3.org" mailing list.

1. Abstract

   The HTTP/1.1 specification (RFC 2068) is silent about various details
   of TCP connection management when using persistent connections. This
   document discusses some of the implementation issues discussed during
   HTTP/1.1's design, and introduces a few new requirements on HTTP/1.1
   implementations learned from implementation experience, not fully
   understood when RFC 2068 was issued. This is an initial draft for
   working group comment, and we expect further drafts.

Gettys & Freier [Page 1]

Internet-Draft HTTP Connection Management March 1997

2. Table of Contents

   1. Abstract ....................................................... 1
   2. Table of Contents .............................................. 2
   3. Key Words ...................................................... 2
   4. Connection Management for Large Scale HTTP Systems ............. 2
   5. Resource Usage (Who is going to pay?) .......................... 2
   6. Go to the Head of the Line ..................................... 6
   7. The Race is On ................................................. 7
   8. Closing Half of the Connection ................................. 8
   9. Capture Effect ................................................. 9
   10. Security Considerations ...................................... 10
   12. References ................................................... 12
   13. Acknowlegements .............................................. 13
   14. Authors' Addresses ........................................... 13

3. Key Words

   The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
   "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this
   document are to be interpreted as described in RFC xxxx. [Bradner]

4. Connection Management for Large Scale HTTP Systems

   Recent development of popular protocols (such as HTTP, LDAP, ...)
   have demonstrated that the standards and engineering communities have
   not yet come to grip with the concept of connection management. For
   instance, HTTP/1.0 [HTTP/1.0] uses a TCP connection for carrying
   exactly one request/response pair. The simplistic beauty of that
   model has much less than optimal behavior.

   This document focuses HTTP/1.1 implementations but the conclusions
   drawn here may be applicable to other protocols as well.

   The HTTP/1.1 Proposed Standard [HTTP/1.1] specification is silent on
   when, or even if, the connection should be closed (implementation
   experience was desired before the specification was frozen on this
   topic). So HTTP has moved from a model that closed the connection
   after every request/response to one that might never close. Neither
   of these two extremes deal with "connection management" in any
   workable sense.

5. Resource Usage (Who is going to pay?)

   The Internet is all about scale: scale of users, scale of servers,

Gettys & Freier [Page 2]

Internet-Draft HTTP Connection Management March 1997

   scale over time, scale of traffic. For many of these attributes,
   clients must be cooperative with servers.

   Clients of a network service are unlikely to communicate with more
   than a few servers (small number of 10s). Considering the power of
   desktop machines of today, maintaining that many idle connections
   does not appear to be overly burdensome, particularly when you
   consider the client is the active party and doesn't really have to
   pay attention to the connection unless it is expecting some response.

   Servers will find connections to be critical resources and will be
   forced to implement some algorithm to shed existing connections to
   make room for new ones. Since this is an area not treated by the
   protocol, one might expect a variety of "interesting" efforts.

   Maintaining an idle connection is almost entirely a local issue.
   However, if that local issue is too burdensome, it can easily become
   a network issue. A server, being passive, must always have a read
   pending on any open connection. Some implementations of the multi-
   wait mechanisms tend to bog down as the number of connections climbs
   in to the hundreds, though operating system implementations can scale
   this into the thousands, tens of thousands, or even beyond. Whether
   server implementations can also scale to so many simultaneous clients
   is likely much less clear than if the operating system can
   theoretically support such use. Implementations might be forced to
   use fairly bizarre mechanisms, which could lead to server
   instability, and then perhaps service outages, which are indeed a
   network issues. And despite any heroic efforts, it will all be to no
   avail. The number of clients that could hold open a connection will
   undoubtedly overwhelm even the most robust of servers over time.

   When this happens, the server will of necessity be forced to close
   connections. The most often considered algorithm is an LRU. The
   success of LRU algorithms in other areas of computer engineering is
   based on locality of reference. I.e., in this case, if LRU is better
   than random, then this is because the "typical" client's behavior is
   predictable based on its recent history. Clients that have made
   requests recently are probably more likely to make them again, than
   clients which have been idle for a while. While we are not sure we
   can point to rigorous proof of this principle, we believe it does
   hold for Web service and client reference patterns are certainly a
   very powerful "clue".

   The client has more information that could be used to drive the
   process. For instance, it does not seem to much to expect that a
   connection be held throughout the loading of a page and all its
   embedded links. It could further sense user sincerity towards the
   page by detecting such events as mouse movement, scrolling, etc., as
   indicators that there is still some interest in pursing the page's
   content, and therefore the chance of accessing subsequent links. But
   if the user has followed a number of links in succession away to a
   different server, it may be likely that the first connection will not

Gettys & Freier [Page 3]

Internet-Draft HTTP Connection Management March 1997

   be used again soon. Whether this is significantly better than LRU is
   an open question, but it is clear that unlikely to be used
   connections should be closed, to free the server resouces involved.
   Server resouces are much more scarce than client resources, and
   clients should be frugal, if the Web is to have good scaling
   properties.

   Authoritative knowledge that it is appropriate to close a connection
   can only come from the user. Unfortunately, that source is not to be
   trusted. First, most users don't know what a connection is, and
   having them indicate it is okay to close it is meaningless. Second, a
   user that does know what a connection is probably inherently greedy.
   Such a user would never surrender the attention that a connection to
   a server implies. Research [Mogul2] does show that most of the
   benefits of persistent connections are gained if connections can be
   held open after last use approximately one minute for the HTTP
   traffic studied; this captures most "click ahead" behavior of a
   user's web browsing.

   For many important services, server resources are critical resources;
   there are many more clients than services. For example, the AltaVista
   search service handles (as of this writing) tens of millions of
   searches per day, for millions of different clients. While it is one
   of the two or three most popular services on the Internet today, it
   is clearly small relative to future services built with Internet
   technology and HTTP. From this perspective, it is clear that clients
   need to cooperate with servers to enable servers to continue to
   scale.

   System resources at a server:

       * Server resources (open files, file system buffers, processes,
         memory for applications, memory for socket buffers for
         connections currently in use (16-64Kbytes each, data base
         locks). In BSD derived TCP implementations, socket buffers are
         only needed on active connections. This usually works because
         it's seldom the case that there is data queued to/from more
         than a small fraction of the open connections.

       * PCB (Protocol control blocks, only ~100-140 bytes; even after a
         connection is closed, you can't free this data structure for a
         significant amount of time, of order minutes. More severe,
         however, is that many inferior TCP implementations have had
         linear or quadratic algorithms relating to the number of PCB's
         to find PCB's when needed.

   These are organized from most expensive, to least.

   Clients should read data from their TCP implementations aggressively,
   for several reasons:

       * TCP implementations will delay acknowledgements if socket

Gettys & Freier [Page 4]

Internet-Draft HTTP Connection Management March 1997

         buffers are not emptied. This will lower TCP performance, and
         cause increased elapsed time for the end user. [Frystyk et.
         al.] while continuing to consume the server's resources.

       * Servers must be able to free the resources held on behalf of
         the client as quickly as possible, so that the server can reuse
         these resources on behalf of others. These are often the
         largest and scarcest server system resource (processes, open
         files, file system buffers, data base locks, etc.)

   When HTTP requests complete (and a connection is idle), an open
   connection still consumes resources some of which are not under the
   server's control:

       * socket buffers (16-64KB both in the operating system, and often
         similar amounts in the server process itself)

       * Protocol Control Blocks (.15 KB/PCB's). (??? Any other data
         structures associated with PCB's?)

   If, for example, an HTTP server had to indefinitely maintain these
   resources, this memory alone for a million clients (and there are
   already HTTP services larger than this scale in existence today)
   using a single connection each would be tens of gigabytes of memory.
   One of the reasons the Web has succeeded is that servers can, and do
   delete connections, and require clients to reestablish connections.

   If connections are destroyed too aggressively (HTTP/1.0 is the
   classic limiting case), other problems ensue.

       * The state of congestion of the network is forgotten [Jacobson].
         Current TCP implementations maintain congestion information on
         a per-connection basis, and when the connection is closed, this
         information is lost. The consequences of this are well known:
         general Internet congestion, and poor user performance

       * Round trip delays and packets to re-establish the connections.
         Since most objects in the Web are very small, of order half the
         packets in the network has been due to just the TCP open and
         close operation.

       * Slow Start lowers initial throughput of the TCP connection

       * PCB's become a performance bottleneck in some TCP
         implementations (and cannot be reused for a XXX timeout after
         the connection has been terminated). The absolute number of
         PCBs in the TIME_WAIT state could be much larger than the
         number in the ESTABLISHED state. Closing connections too
         quickly can actually consume more memory than closing them
         slowly, because all PCBs consume memory and idle socket buffers
         do not.

Gettys & Freier [Page 5]

Internet-Draft HTTP Connection Management March 1997

   From these two extreme examples, it is obvious that connection
   management becomes a central issue for both clients and servers.

   Clearly, benefits of persistent connections will be lost if clients
   open many connections simultaneously. RFC2068 therefore specifies no
   more than 2 connections from a client to a server should be open at
   any one time, or 2N connections (where N is the number of clients a
   proxy is serving) for proxies. Frystyk et. al. have shown that
   roughly twice the performance of HTTP/1.0 using four to six
   connections can be reached using HTTP/1.1 over a single TCP
   connection using HTTP/1.1, even over a LAN, once combined with
   compression of the HTML documents [Frystyk].

6. Go to the Head of the Line

   The HTTP/1.1 specification requires that proxies use no more than 2N
   connections, where N is the number of client connections being
   served. Mogul has shown that persistent connections are a "good
   thing", and Frystyk et. al. show data that significant (a factor of
   2-8) savings in number of packets transmitted result by using
   persistent connections.

   If fewer connections are better, then, why does HTTP/1.1 permit
   proxies to establish more than the absolute minimum of connections?
   In the interests of brevity, the HTTP/1.1 specification is silent on
   some of the motivations for some requirements of the specification.
   At the time HTTP/1.1 was specified, we realized that if a proxy
   server attempted to aggregate requests from multiple client
   connections onto a single TCP connection, a proxy would become
   vulnerable to the "head of line" blocking problem. If Client A, for
   example, asks for 10 megabytes of data (or asked for a dynamicly
   generated document of unlimited length), then if a proxy combined
   that request with requests from another Client B, Client B would
   never get its request processed. This would be a very "bad thing",
   and so the HTTP/1.1 specification allows proxies to scale up their
   connection use in proportion to incoming connections. This will also
   result in proxy servers getting roughly fair allocation of bandwidth
   from the Internet proportional to the number of clients.

   Since the original HTTP/1.1 design discussions, we realized that
   there is a second, closely related denial of service security arises
   if proxies attempt to use the same TCPconnection for multiple
   clients. An attacker could note that a particular URL of a server
   that they wished to attack was either very large, very slow (script
   based), or never returned data. By making requests for that URL, the
   attacker could easily block other clients from using that server
   entirely, due to head of line blocking, so again, simultaneously
   multiplexing requests from different clients would be very bad, and
   therefore implementations MUST not attempt such multipexing.

   In other words, head-of-line blocking couples the fates of what
   should be independent interactions, which allows for both denial-of-

Gettys & Freier [Page 6]

Internet-Draft HTTP Connection Management March 1997

   service attacks, and for accidental synchronization.

   Here is another example of head-of-line blocking: imagine clients A
   and B are connected to proxy P1, which is connected to firewall proxy
   P2, which is connected to the Internet. If P1 only has one connection
   to P2, and A attempts to connect (via P1 and P2) to a dead server on
   the Internet, all of B's operations are blocked until the connection
   attempt from P2 to the dead server times out. This is not a good
   situation.

   Note that serial reuse of a TCP connection does not have these
   considerations: a proxy might first establish a connection to an
   origin server for Client A, and possibly leave the connection open
   after Client A finishes and closes

   its connection, and then use the same connection for Client B, and so
   on. As in normal clients, such a proxy should close idle
   connections.

   Future HTTP evolution also dictates that simultaneous multiplexing of
   clients over a connection should be prohibited. A number of schemes
   for compactly encoding HTTP rely on associating client state with a
   connection, which HTTP 1.X does not currently do. If proxies do such
   multiplexing, then such designs will be much harder to implement.

7. The Race is On

   Deleting a connection without authoritative knowledge that it will
   not be soon reused is a fundamental race that is part of any timeout
   mechanism. Depending on how the decision is made will determine the
   penalties imposed.

   It is intuitively (and most certainly empirically) less expensive for
   the active (client) partner to close a connection than the server.
   This is due in most part to the natural flow of events. For instance,
   a server closing a connection cannot know that the client might at
   that very moment be sending a request. The new request and the close
   message can pass by in the night simply because the server and the
   client are separated by a network. That type of failure is a network
   issue. The code of both the client and the server must to be able to
   deal with such failures, but they should not have to deal with it
   efficiently. A client closing a connection, on the other hand, will
   at least be assured that any such race conditions are mostly local
   issues. The flow will be natural, assuming one treats closing as a
   natural event. To paraphrase Butler Lampson's 1983 paper on system
   design, "The events that happen normally must be efficient. The
   exceptional need to make progress." [Lampson]

   Having the client closing the connection will decrease the
   probability of the client having to do automatic connection recovery
   of a pipeline caused by a premature close on server side. From an
   client implementation point of view this is advantageous as automatic

Gettys & Freier [Page 7]

Internet-Draft HTTP Connection Management March 1997

   connection recovery of a pipeline is significantly more complicated
   than closing an idle connection. In HTTP, however, servers are free
   to close connections any time, and this observation does not help,
   but may simplify other protocols. It will, however, reduce the number
   of TCP resets observed, and make the exceptional case exceptional,
   and avoid a TCP window full of requests being transmitted under some
   circumstances.

   On the one hand, it is a specific fact about TCP that if the client
   closes the connection, the server does not have to keep the TIME_WAIT
   entry lying around. This is goodness.

   On the other hand, if the server has the resources to keep the
   connection open, then the client shouldn't close it unless there is
   little chance that the client will use the server again soon, since
   closing & then reopening adds computational overhead to the server.
   So allowing the server to take the lead in closing connections does
   have some benefits.

   A further observation is that congestion state of the network varies
   with time, so the benefits of the congestion state being maintained
   by TCP diminishes the longer a connection is idle.

   This discussion also shows that a client should close idle
   connections before the server does. Currently in the HTTP standard
   there is no way for a server to provide such a "hint" to the client,
   and there should be a mechanism. This memo solicits other opinions on
   this topic.

8. Closing Half of the Connection

   In simple request/response protocols (e.g. HTTP/1.0), a server can go
   ahead and close both recieve and transmit sides of its connection
   simultaneously whenever it needs to. A pipelined or streaming
   protocol (e.g. HTTP/1.1) connection, is more complex [Frystyk et.
   al.], and an implementation which does so can create major problems.

   The scenario is as follows: an HTTP/1.1 client talking to a HTTP/1.1
   server starts pipelining a batch of requests, for example 15 on an
   open TCP connection. The server decides that it will not serve more
   than 5 requests per connection and closes the TCP connection in both
   directions after it successfully has served the first five requests.
   The remaining 10 requests that are already sent from the client will
   along with client generated TCP ACK packets arrive on a closed port
   on the server. This "extra" data causes the server's TCP to issue a
   reset which makes the client TCP stack pass the last ACK'ed packet to
   the client application and discard all other packets. This means that
   HTTP responses that are either being received or already have been
   received successfully but haven't been ACK'ed will be dropped by the
   client TCP. In this situation the client does not have any means of
   finding out which HTTP messages were successful or even why the
   server closed the connection. The server may have generated a

Gettys & Freier [Page 8]

Internet-Draft HTTP Connection Management March 1997

   "Connection: Close" header in the 5th response but the header may
   have been lost due to the TCP reset. Servers must therefore close
   each half of the connection independently.

9. Capture Effect

   One of the beauties of the simple single connection for each
   request/response pair is that it did not favor an existing client
   over another. In general, this natural rotation made for a fairer
   offering of the overall service, albeit a bit heavy handed. Our
   expectation is that when protocols with persistent connections get
   heavily deployed, that aspect of fairness will not exist. Without
   some moderately complex history, it might be that only the first 1000
   clients will ever be able to access a server (providing that your
   server can handle 1000 connections).

   There needs to be some policy indicating when it is appropriate to
   close connections. Such a policy should favor having the client be
   the party to initiate the closure, but must provide some manner in
   which the server can protect itself from misbehaving clients. Servers
   can control greedy clients in HTTP/1.1 by use of the 503 (Service
   Unavailable) response code in concert with the Retry-After response-
   header field, or by not reading further requests from that client, at
   the cost of temporarily occupying the connection. As long as the
   server can afford to keep the connection open, it can delay a "greedy
   client" by simply closing the TCP receive window. As soon as it
   drops the connection, it has no way to distinguish this client from
   any other. Either of these techniques may in fact be preferable to
   closing the client's connection; the client might just immediately
   reopen the connection, and you are unlikely to know if it is the same
   greedy client.

   Implementation complexity will need to be balanced against scheduling
   overhead. A number of possible server scheduling algorithms exist,
   with different costs and benefits. The implementation experience of
   one of us (jg) with the X Window System [Gettys et. al.] may be of
   use to those implementing Web server schedulers.

       * Strict round robin scheduling: a operating system select or
         poll operation is executed for each request processed, and each
         request is handled in turn (across connections). Since select
         is executed frequently, new connections get a good chance of
         service sooner rather than later. Some algorithm must be chosen
         to avoid capture effect if the server is loaded. This is most
         fair, and approximates current behavior. The disadvantage is,
         however, a (relatively expensive) system call / request, which
         will likely become too expensive as Web servers become
         carefully optimized after HTTP/1.1 is fully implemented.

       * Modified round robin scheduling: a operating system select or
         poll operation is executed. Any new connections are
         established, and for each connection showing data available,

Gettys & Freier [Page 9]

Internet-Draft HTTP Connection Management March 1997

         all available requests are read into buffers for later
         execution. Then all requests are processed, round robin between
         buffers. Some algorithm must be chosen to avoid capture effect
         if the server is loaded. This eliminates the system call per
         operation. This is quite efficient, and still reasonably
         fairly apportions server capabilities.

       * Some servers are likely to be multithreaded, possibly with a
         thread per connection. These servers will have to have some
         mechanism to share state so that no client can forever capture
         a connection on a busy server.

   A final note: indefinite round robin scheduling may not in fact be
   the most desirable algorithm, due to the timesharing fallacy. If a
   connection makes progress more slowly than possible, not only will
   the client (the end user) observe poorer performance, but the
   connection (and the considerable system overhead each one represents)
   will be open longer, and more connections and server resources will
   be required as a result.

   At some point, large, loaded servers will have to choose a connection
   to close; research [Padmanabhan and Mogul] shows that LRU may be as
   good as more complex algorithms for choosing which to close.

   Further experimentation with HTTP/1.1 servers will be required to
   understand the most useful scheduling and connection management
   algorithms.

10. Security Considerations

   Most HTTP related security considerations are discussed in RFC2068.
   This document identifies a further security concern: proxy
   implementations that simultaneously multiplex requests from multiple
   clients over a TCP connection are vulnerable to a form of denial of
   service attacks, due to head of line blocking problems, as discussed
   further above.

   The capture effect discussed above also presents opportunities for
   denial of service attacks.

11. Requirements on HTTP/1.1 Implementations

   Here are some simple observations and requirements from the above
   discussion.

       * clients and proxies SHOULD close idle connections. Most of the
         benefits of an open connection diminish the longer the
         connection is idle: the congestion state of the network is a
         dynamic and changing phenomena [Paxson]. The client, better
         than a server, knows when it is likely not to revisit a site.
         By monitoring user activity, a client can make reasonable
         guesses as to when a connection needs closing. Research has

Gettys & Freier [Page 10]

Internet-Draft HTTP Connection Management March 1997

         shown [Mogul2] shows that most of the benefits of a persistent
         connection are likely to occur within approximately
         60 seconds. Further research in this area is needed. On the
         client side, define a connection as "idle" if it meets at least
         one of these two criteria:

             * no user-interface input events during the last 60 seconds
               (parameter value shouldn't be defined too precisely)

             * user has explicitly selected a URL from a different
               server. Don't switch just because inlined images are from
               somewhere else! Even in this case, dally for some seconds
               (e.g., 10) in case the user hits the "back" button.
         On the server side, use a timeout that is adapted based on
         resource constraints: short timeout during overload, long
         timeout during underload. Memory, not CPU cycles, is likely to
         be the controlling resource in a well-implemented system.

       * servers SHOULD implement some mechanism to avoid the capture
         effect.

       * proxies MUST use independent TCPconnections to origin or futher
         proxy servers for different client connections, both to avoid
         head of line blocking between clients, and to avoid the denial
         of service attacks that implementations that attempt to
         multiplex multiple clients over the same connection would be
         open to.

       * proxies MAY serially reuse connections for multiple clients.

       * servers MUST properly close incoming and outgoing halves of TCP
         connections independently.

       * clients SHOULD close connections before servers when possible.
         Currently, HTTP has no "standard" way to indicate idle time
         behavior to clients, though we note that the Apache HTTP/1.1
         implementation advertizes this information using the Keep-Alive
         header if Keep-Alive is requested. We note, however, that Keep-
         Alive is NOT currently part of the HTTP standard, and that the
         working group may need to consider providing this "hint" to
         clients in the future of the standard by this or other means
         not currently specified in this initial draft.

Gettys & Freier [Page 11]

Internet-Draft HTTP Connection Management March 1997

12. References

   [Apache]
      The Apache Authors, The Apache Web Server is distributed by The
      Apache Group.

   [Bradner]
      S. Bradner, "Keywords for use in RFCs to Indicate Requirement
      Levels", RFC XXXX

   [Frystyk]
      Henrik Frystyk Nielsen, "The Effect of HTML Compression on a LAN
      ", W3C. URL:
      http://www.w3.org/pub/WWW/Protocols/HTTP/Performance/Compression/LAN.html

   [Frystyk et. al]
      Henrik Frystyk Nielsen, Jim Gettys, Anselm Baird-Smith, Eric
      Prud'hommeaux, W3C, H&aring;kon Wium Lie, Chris Lilley, W3C,
      "Network Performance Effects of HTTP/1.1, CSS1, and PNG". W3C
      Note, February, 1997. See URL:
      http://www.w3.org/pub/WWW/Protocols/HTTP/Performance/ for this and
      other HTTP/1.1 performance information.

   [Gettys et. al.]
      Gettys, J., P.L. Karlton, and S. McGregor, " The X Window System,
      Version 11.'' Software Practice and Experience Volume 20, Issue
      No. S2, 1990 ISSN 0038-0644.

   [HTTP/1.0]
      T. Berners-Lee, R. Fielding, H. Frystyk. "Informational RFC 1945
      - Hypertext Transfer Protocol -- HTTP/1.0," MIT/LCS, UC Irvine,
      May 1996

   [HTTP/1.1]
      R. Fielding, J. Gettys, J.C. Mogul, H. Frystyk, T. Berners-Lee,
      "RFC 2068 - Hypertext Transfer Protocol -- HTTP/1.1," UC Irvine,
      Digital Equipment Corporation, MIT

   [Jacobson]
      Van Jacobson. "Congestion Avoidance and Control." In Proc. SIGCOMM
      '88 Symposium on Communications Architectures and Protocols, pages
      314-329. Stanford, CA, August, 1988.

   [Lampson]
      B. Lampson, "Hints for Computer System Design", 9th ACM SOSP, Oct.
      1983, pp. 33-48.

   [Mogul]
      Jeffrey C. Mogul. "The Case for Persistent-Connection HTTP." In
      Proc. SIGCOMM '95 Symposium on Communications Architectures and
      Protocols, pages 299-313. Cambridge, MA, August, 1995.

Gettys & Freier [Page 12]

Internet-Draft HTTP Connection Management March 1997

   [Mogul2]
      Jeffrey C. Mogul. "The Case for Persistent-Connection HTTP".
      Research Report 95/4, Digital Equipment Corporation Western
      Research Laboratory, May, 1995. URL:
      http://www.research.digital.com/wrl/techreports/abstracts/95.4.html

   [Padmanabhan and Mogul]
      Venkata N. Padmanabhan and Jeffrey C. Mogul. Improving HTTP
      Latency. In Proc. 2nd International WWW Conf. '94: Mosaic and the
      Web, pages 995-1005. Chicago, IL, October, 1994. URL:
      http://www.ncsa.uiuc.edu/SDG/IT94/Proceedings/DDay/mogul/HTTPLatency.html

   [Padmanabhan & Mogul]
       V.N. Padmanabhan, J. Mogul, "Improving HTTP Latency", Computer
      Networks and ISDN Systems, v.28, pp. 25-35, Dec. 1995. Slightly
      revised version of paper in Proc. 2nd International WWW Conference
      '94: Mosaic and the Web, Oct. 1994

   [Paxson]

      Vern Paxson, "End-to-end Routing Behavior in the Internet" ACM
      SIGCOMM '96, August 1996, Stanford, CA.

13. Acknowlegements

   Our thanks to Henrik Frystyk Nielsen for comments on the first draft
   of this document.

14. Authors' Addresses

   Jim Gettys
   W3 Consortium
   MIT Laboratory for Computer Science
   545 Technology Square
   Cambridge, MA 02139, USA
   Fax: +1 (617) 258 8682
   Email: jg@w3.org

   Alan Freier
   Netscape Communications Corporation
   Netscape Communications
   501 East Middlefield Rd.
   Mountain View, CA 94043
   Email: freier@netscape.com

Gettys & Freier [Page 13]

-- 
miguel a.l. paraz  <map@iphil.net>                              +63-2-893-0850
iphil communications, makati city, philippines          <http://www.iphil.net> 
Received on Tue Jul 29 2003 - 13:15:40 MDT

This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 16:11:15 MST