[pilc] A few comments on LINK version 11

From: Spencer Dawkins (spencer_dawkins@yahoo.com)
Date: Tue Jun 11 2002 - 13:50:50 EDT

  • Next message: Joe Touch: "Re: [pilc] A few comments on LINK version 11"

    Dear All,

    I've been through LINK again. When PILC was formed,
    the expectation was that
    this document would be our primary output, with the
    (aready-approved) BCPs
    helping to bandaid TCP over already-deployed
    subnetwork technologies. I
    believe that Phil and Mark have met this expectation,
    with help from a number
    of others.

    I do have a few comments on version 11...

    Spencer

    ------------

    1 Introduction and Overview

    I think the following is too polite!

       Because IP is so simple, consideration 2 is more of
    an issue than
       consideration 1. I.e., subnetwork designers make
    many more errors of
       commission than errors of omission. But certain
    enhanced Internet
       features, such as multicasting and
    quality-of-service, benefit
       significantly from support from the underlying
    subnetworks beyond
       that necessary to carry "traditional" unicast,
    best-effort IP.

    I propose

       IP transport requirements are so minimal that
    consideration 2 is more
       of an issue than consideration 1. Subnetwork
    designers make many more
       errors of commission than errors of omission.

       o A frequent error of commission is designing a
    subnetwork that provides
         a high level of subnetwork-level reliability for
    IP datagrams, at a
         corresponding high cost in latency and jitter -
    even though higher-level
         protocols can't assume this level of reliability
    for all IP paths and
         must duplicate this reliability if it's needed.

       o When subnetwork designers do make errors of
    omission, it's usually
         by not providing subnetwork-level support for
    multicast transport
         and quality-of-service mechanisms - or anything
    else beyond
         "traditional" unicast, best-effort IP transport.

    Would it make sense to point to RFC 3135 at the end of
    the following paragraph?

       However, partial duplication of functionality in a
    lower layer can
       *sometimes* be justified by performance, security
    or availability
       considerations. Examples include link-layer
    retransmission to improve
       the performance of an unusually lossy channel,
    e.g., mobile radio;
       link level encryption intended to thwart traffic
    analysis; and
       redundant transmission links to improve
    availability, increase
       throughput, or to guarantee performance for certain
    classes of
       traffic. Duplication of protocol function should
    be done only with
       an understanding of system level implications,
    including possible
       interactions with higher-layer mechanisms.

    2 Maximum Transmission Units (MTUs) and IP
    Fragmentation

    I had too may "it is" in this paragraph to parse
    quickly:

       If optional header compression [RFC1144] [RFC2507]
    [RFC2508]
       [RFC3095] is used, however, it is required that the
    link framing
       indicate frame length as it is needed for the
    reconstruction of the
       original header.

    I propose that we use more nouns, as follows:

       If optional header compression [RFC1144] [RFC2507]
    [RFC2508]
       [RFC3095] is used, however, link framing mechanisms
    are required to
       indicate frame length as this information is needed
    to reconstruct
       the original header.

    Do we still think P-MTU black holes are a sufficent
    problem that we
    will still include this problem on the following list
    of justifications
    for support of large MTU sizes? (If so, should P-MTU
    discovery still be
    on the 2.5G3G list of recommendations? But that's not
    a LINK question...)

       The Path MTU Discovery procedure, the persistence
    of path MTU black
       holes, and the deletion of router fragmentation in
    IPv6 reflects a
       consensus of the Internet technical community that
    router
       fragmentation is best avoided. This requires that
    subnetworks support
       MTUs that are "reasonably" large. The smallest MTU
    permitted in IPv4
       by [RFC791] is 68 bytes, but such a small value
    would clearly be
       inefficient. Because IPv6 omits fragmentation by
    routers, [RFC 2460]
       specifies a larger minimum MTU of 1280 bytes. Any
    subnetwork with an
       internal packet payload smaller than 1280 bytes
    must implement a
       mechanism that performs fragmentation/reassembly of
    IP packets
       to/from subnetwork frames if it is to support IPv6.

    Is there an interaction here between internal
    fragmentation and
    insertion of a smaller/higher priority packet during
    transmission
    of a larger/lower priority packet, for reasons
    described in 2.1, and using
    the multiplexing mechanism described in 2.1?

       If a subnetwork cannot directly support a
    "reasonable" MTU with
       native framing mechanisms, it should internally
    fragment. That is, it
       should transparently break IP packets into internal
    data elements and
       reassemble them at the other end of the subnetwork.

       This leaves the question of what is a "reasonable"
    MTU. Ethernet (10
       and 100 Mb/s) has a MTU of 1500 bytes, and because
    of its ubiquity

    Isn't 1500 bytes the limit for all 802.3s (including 1
    and 10 Gb/s)?

       few Internet paths have MTUs larger than this
    value. This severely
       limits the utility of larger MTUs provided by other
    subnetworks.
       Meanwhile larger MTUs are increasingly desirable on
    high speed
       subnetworks to reduce the per-packet processing
    overhead in host
       computers, and implementers are encouraged to
    provide them even
       though they may not be usable when Ethernet is also
    in the path.

       Various "tunneling" schemes, such as IP Security
    [RFC2406] treat IP
       as a subnetwork for IP. Since tunneling adds
    header overhead, it can

    Doesn't insertion of MPLS headers have the same
    effect?

       trigger fragmentation even when the same physical
    subnetworks (e.g.,
       Ethernet) are used on both sides of the IP router.
    Tunneling has made
       it more difficult to avoid router fragmentation and
    has increased the
       incidence of path MTU black holes. Larger
    subnetwork MTUs may help to
       alleviate this problem.

    I think it's also reasonable to point out that P-MTU
    discovery uses a
    table of "likely" MTU sizes that doesn't map very well
    to what used to
    be full-sized MSSes that are being tunneled - the
    fallback to the first
    working MTU in the table may be quite a bit smaller
    than the un-tunneled
    P-MTU... I wonder if this table will ever be
    updated...

     2.1 Choosing the MTU in Slow Networks

       In slow networks, the largest possible packet may
    take a considerable
       time to send. Interactive response time should not
    exceed the well-

    Can (should) we say "Round-trip times for interactive
    responses should
    not exceed"?

    Also, can we point to RFC 1144 as a source for the
    100-200 ms requirement
    for interactive response?

       known human factors limit of 100 to 200 ms. This
    includes all sources
       of delay: electromagnetic propagation delay,
    queuing delay, and the
       store-and-forward time, i.e,. the time to transmit
    a packet at link
       speed.

    3 Framing on Connection-Oriented Subnetworks

    (no comments in this section)

    4 Connection-Oriented Subnetworks

       Because Internet traffic is typically bursty and
    transaction-
       oriented, it is often difficult to pick an optimal
    idle timeout. If
       the timeout is too short, subnetwork connections
    are opened and
       closed rapidly, possibly over-stressing its call
    management system
       (especially if was designed for voice traffic
    holding times). If the
       timeout is too long, subnetwork connections are
    idle much of the
       time, wasting any resources dedicated to them by
    the subnetwork.

    Can we also add the following:

       Subscribers may also be charged differently
    depending on whether
       a subnetwork leaves idle connections in place using
    a
       relatively-long idle timer or tears down idle
    connections with
       little delay and reestablishes these connections
    after a short
       period of time. If it makes sense to do so,
    subscribers should be
       given control over this behavior (for example, PPP
    implementations
       might allow subscribers to control an idle timer
    for locally-
       initiated dialup connections.

    Should be "is used TO support IP" in the following
    sentence.

       In any event, if an SNDCF that opens and closes
    subnet connections is
       used support IP, care should be taken to make sure
    that call
       processing in the subnet can keep up with
    relatively short holding
       times.

    5 Broadcasting and Discovery

       Subnetworks fall into two categories:
    point-to-point and shared. A
       point-to-point subnet has exactly two endpoint
    components (hosts or
       routers); a shared link has more than two, using
    either an inherent
       broadcast media (e.g., Ethernet, radio) or on a
    switching layer
       hidden from the network layer (switched Ethernet,
    Myrinet [MYR],
       ATM). Switched subnetworks handle broadcast by
    copying broadcast

    Could this sentence say "by forwarding broadcast
    packets on all other
    interfaces to ensure..."?

       packets to ensure each end system receives a copy
    of each packet.

    Could we insert "Centralized databases also insert new
    failure points and
    scaling hot-spots into the network." before the last
    sentence in the
    following paragraph?

       The lack of broadcast can impede the performance of
    these protocols,
       or render them inoperable (e.g. DHCP). ARP-like
    link address lookup
       can be provided by a centralized database, but at
    the expense of
       potentially higher response latency and the need
    for nodes to have
       explicit knowledge of the ARP server address.
    Shared links should
       support native, link-layer subnet broadcast.

    6 Multicasting

       Receivers also need to be designed to accept
    packets addressed to
       some number of multicast addresses in addition to
    the unicast packets
       specifically addressed to them. How many multicast
    addresses need to
       be supported by a host depends on the requirements
    of the associated
       host; at least several dozen will meet most current
    needs.

    The last phrase doesn't parse well for me. Is it
    saying "few hosts
    must accept multicast packets from more than a few
    dozen multicast
    addresses"?

       On low-speed networks the multicast address
    recognition function may
       be readily implemented in host software, but on
    high speed networks
       it should be implemented in subnetwork hardware.
    This hardware need
       not be complete; for example, many Ethernet
    interfaces implement a
       "hashing" function that passes all of the multicast
    (and unicast)
       traffic to which the associated host subscribes,
    plus some small
       fraction of multicast traffic to which the host
    does not subscribe.
       Host/router software then has to discard the
    unwanted packets that
       pass the hardware filter.

       There does not need to be a one-to-one mapping
    between subnetwork
       multicast address and IP multicast address. An
    address overlap may
       significantly degrade the filtering capability of a
    receiver's
       hardware multicast address filter. A subnetwork
    supporting only

    Is this sentence saying "If more than one IP multicast
    address makes into
    a subnetwork multicast address, this many-to-one
    mapping may significantly
    degrade the filtering capacity of a receiver's
    hardware multicast address
    filter", or something else? "Address overlap" is too
    ambiguous for me...

       broadcast should use this service for multicast and
    must rely on
       software filtering.

    7 Bandwidth on Demand (BoD) Subnets

       Long delay BoD subnets pose problems similar to
    connection oriented
       networks in anticipating traffic. While connection
    oriented subnets
       hold idle channels open expecting new data to
    arrive, BoD subnets
       request channel access based on buffer occupancy
    (or expected buffer
       occupancy) on the sending port. Poor performance
    will likely result
       if the sender does not anticipate additional
    traffic arriving at that

    I'm confused here. Is this sentence saying "Poor
    performance will likely
    resuly if the sending host does not queue additional
    outgoing traffic
    while the sender is waiting for its transmission
    request to be granted"?
    If so, I agree. If not, can we talk? I may not
    disagree, I just don't
    understand the point being made.

       port during the time it takes to grant a
    transmission request. It is
       recommended that the algorithm have the capability
    to extend a hold
       on the channel for data that has arrived after the
    original request
       was generated (this may done by piggybacking new
    requests on user
       data).

       There are a wide variety of BoD protocols
    available. However, there
       has been relatively little comprehensive research
    on the interactions
       between the BoD mechanisms and Internet protocol
    performance.
       Research on some specific mechanisms is available
    (e.g., [AR02]).
       One item that has been studied is TCP's
    retransmission timer [KY02].
       BoD systems can cause spurious timeouts when
    adjusting from a
       relatively high data rate to a relatively low data
    rate. In this
       case, TCP's transmitted data takes longer to get
    through the network
       than predicted by the retransmission timeout (RTO)
    and therefore the
       TCP sender is prone to resending a segment
    prematurely.

    Could we also include the observation that delays to
    acquire a shared
    control channel may also be reflected as sudden and
    unpredictable "spikes"
    in apparent round-trip times, making RTO values
    "spike" unpredictably
    as well? I'm trying to say that it's not just the
    change in available
    bandwidth, but delays encountered while changing
    available bandwidth
    as well...

    8 Reliability and Error Control

       In the Internet architecture, the ultimate
    responsibility for error
       recovery is at the end points. The Internet may
    occasionally drop,
       corrupt, duplicate or reorder packets, and the
    transport protocol
       (e.g., TCP) or application (e.g., if UDP is used as
    the transport
       protocol) must recover from these errors on an
    end-to-end basis.

    I'm not sure this is strictly true - I'm thinking of
    RTP over UDP,
    where the RTP sender may use RTCP indications of loss
    to lower its sending
    rates. Is this "recovering" in the "reliability and
    error control"
    sense? I think it's not "must recover", but "recovers
    if necessary".

       Error recovery in the subnetwork is therefore
    justified only to the
       extent that it can enhance overall performance. It
    is important to
       recognize that a subnetwork can go too far in
    attempting to provide
       error recovery services in the Internet
    environment. Subnet
       reliability should be "lightweight", i.e., it only
    has to be "good
       enough", *not* perfect.

    Can we add something like:

    Applications have no way of accepting a lower level of
    reliability
    (perhaps as a tradeoff for lower latency) than the
    subnetwork provides.
    "Reliability" is a good thing until it prevents users
    from using an
    application over a too-reliable subnetwork.

       In this section we discuss how to analyze
    characteristics of a
       subnetwork to determine what is "good enough". The
    discussion below
       focuses on TCP, which is the most widely used
    transport protocol in
       the Internet. It is widely believed (and is a
    stated goal within the
       IETF) that non-TCP transport protocols should
    attempt to be "TCP-
       friendly" and have many of the same performance
    characteristics.

    Is this true? I thought the point of SCTP and DCP was
    have "safe"
    behavior when facing congestion, but to vary other
    dimensions of
    performance.

       Thus, the discussion below should be applicable
    even to portions of
       the Internet where TCP may not be the predominant
    protocol.

    I think the discussion is applicable, but not just
    because all transport
    protocols have the same performance characteristics as
    TCP.

     8.1 TCP vs Link-Layer Retransmission

       The use of ECC to detect transmission errors so
    that retransmissions
       (hopefully without errors) can be requested is
    widely known as "ARQ"
       (Automatic Repeat Request).

    Can we make it "Automatic Repeat reQuest", to match
    the acronym?

       This inter-layer "competition" might lead to the
    following wasteful
       situation. When the link layer retransmits (parts
    of) a packet, the
       link latency momentarily increases. Since TCP bases
    its
       retransmission timeout on prior measurements of
    end-to-end latency,
       including that of the link in question, this sudden
    increase in
       latency may trigger an unnecessary retransmission
    by TCP of a packet
       that the link layer is still retransmitting. Such
    spurious end-to-
       end retransmissions generate unnecessary load and
    reduce end-to-end
       throughput. One may even have multiple copies of
    the same packet in
       the same link queue at the same time. In general,
    one could say the
       competing error recovery is caused by an inner
    control loop (link-
       layer error recovery) reacting to the same signal
    as an outer control
       loop (end- to-end error recovery) without any
    coordination between
       the loops. Note that this is solely an efficiency
    issue; TCP
       continues to provide reliable end-to-end delivery
    over such links.

    Can we append ", although this delivery is being
    slowed by unnecessary
    link-layer transmissions"?

     8.2 Recovery from Subnetwork Outages

       Under these circumstances TCP will, after each
    unsuccessful
       retransmission, wait even longer before trying
    again; this is its
       "exponential back-off" algorithm. Furthermore, TCP
    will not discover
       that the subnetwork outage has ended until its next
    retransmission
       attempt. If TCP has backed off, this may take some
    time. This can

    Can we say "this may take a number of minutes."? I'm
    still talking to
    people who find these long timer backoffs surprising.

       lead to extremely poor TCP performance over such
    subnetworks.

       It is therefore highly desirable that a subnetwork
    subject to outages
       not silently discard packets during an outage.
    Ideally, it should
       define an interface to the next higher layer (i.e.,
    IP) that allows
       it to refuse packets during an outage, and to
    automatically ask IP
       for new packets when it is again able to deliver
    them. If it cannot
       do this, then the subnetwork should hold onto at
    least some of the

    We're saying "at least some" in this paragraph and
    "only a single packet
    per TCP connection, including ACKs" in the next
    paragraph. Can we make
    one recommendation?

       packets it accepts during an outage and attempt to
    deliver them when
       the subnetwork comes back up. When packets are
    discarded, IP should

    This is "When outbound packets are discarded,", right?

       be notified so that the appropriate ICMP messages
    can be sent.

       Only a single packet per TCP connection, including
    ACKs, need be held
       in this way to cause the TCP sender to recover from
    the additional
       losses once the flow resumes [ARQ-DRAFT].

    (See previous note about the right number of packets
    to hang onto during
    outages.)

    Is there a maximum period of time that a subnetwork
    should hang onto a
    packet? I don't feel good about "longer than 2 * MSL"
    - would sending
    packets that are older than this cause problems?

     8.3 CRCs, Checksums and Error Detection

       The TCP [RFC793], UDP [RFC768], ICMP, and IPv4
    [RFC791] protocols all
       use the same simple 16-bit 1's complement checksum
    algorithm to
       detect corrupted packets. The IPv4 checksum
    protects only the IPv4
       header, while the TCP, ICMP, and UDP checksums
    provide end-to-end
       error detection for both the transport pseudo
    header (including
       network and transport layer information) and the
    transport payload
       data. Protection of the data is optional for
    applications using UDP
       [RFC768].

    Do we think UDP checksums are "optional but
    recommended", based on the
    NFS war stories, or is this going too far? My
    understanding is that
    UDP checksums are enabled host-wide on most
    implementations - would
    this make turning UDP checksums off less desirable,
    because errored
    packets might be delivered to any running application
    on the host?

       One way to provide additional protection for an
    IPv4 or IPv6 header
       is by the authentication and packet integrity
    services of the IP
       Security (IPSEC) protocol [RFC2401]. However, this
    may not be a
       choice available to the subnetwork designer.

    Is this just saying IPSEC AH is an IP mechanism, or is
    it saying more?

     8.4 How TCP Works

       TCP uses sequence numbering and acknowledgments
    (ACKs) on an end-to-
       end basis to provide reliable, sequenced, once-only
    delivery. TCP
       ACKs are cumulative, i.e., each implicitly ACKs
    every segment
       received so far. If an ACK is not received, the
    acknowledgment value

    Isn't this "implicitly ACKs all previously
    unacknowledged data in the
    sender's outgoing window"?

       carried in the cumulative packet will cease to
    advance.

       Since the most common cause of packet loss is
    congestion, TCP treats
       packet loss as an Internet congestion indicator.
    This happens
       automatically, and the subnetwork need not know
    anything about IP or
       TCP. It simply drops packets whenever it must,
    though some packet-
       dropping strategies (e.g., RED) are more fair to
    competing flows than
       others.

    Is this a fairness issue? I might have said "reduce
    the effect of
    congestion on competing flows, compared to others
    (tail drop)."

    The following seems to ignore slow start:

       TCP recovers from packet losses in two different
    ways. The most
       important is the retransmission timeout. If an ACK
    fails to arrive
       after a certain period of time, TCP retransmits the
    oldest unacked
       packet. Taking this as a hint that the network is
    congested, TCP
       waits for the retransmission to be ACKed before it
    continues, and it
       gradually increases the number of packets in flight
    as long as a
       timeout does not occur again.

    Restart behavior is more correctly described in the
    next paragraph -
    can we drop the last clause of the previous paragraph
    ("... and it
    gradually ...")? Then we go straight into:

       A retransmission timeout can impose a significant
    performance
       penalty, as the sender is idle during the timeout
    interval and
       restarts with a congestion window of 1 following
    the timeout. To
       allow faster recovery from the occasional lost
    packet in a bulk
       transfer, an alternate scheme known as "fast
    recovery" was introduced
       [RFC2581] [RFC2582] [RFC2914] [TCPF98].

    This discussion of fast recovery seems so loose that
    it's actually about
    fast retransmit:

       Fast recovery relies on the fact that when a single
    packet is lost in

    "Fast retransmit relies on the fact that even when one
    or more oackets
    are lost in a bulk transfer, the receiver may continue
    to return ACKs
    to ..."

       a bulk transfer, the receiver continues to return
    ACKs to subsequent
       data packets that do not actually acknowledge any
    newly-received
       data. These are known as "duplicate
    acknowledgments" or "dupacks".
       The sending TCP can use dupacks as a hint that a
    packet has been lost
       and retransmit it without waiting for a timeout.
    Dupacks effectively
       constitute a negative acknowledgment (NAK) for the
    packet sequence
       number in the acknowledgment field. TCP waits
    until a certain number
       of dupacks (currently 3) are seen prior to assuming
    a loss has
       occurred; this helps avoid an unnecessary
    retransmission during out-
       of-sequence delivery.

       A new technique called "Explicit Congestion
    Notification" (ECN)
       [RFC3168] allows routers to directly signal
    congestion to hosts
       without dropping packets. This is done by setting
    a bit in the IP
       header. Since this is currently an optional
    behavior (and, longer
       term, there will always be the possibility of
    congestion in portions
       of the network which don't support ECN), the lack
    of an ECN bit must
       NEVER be interpreted as a lack of congestion.
    Thus, for the

    "lack of congestion - especially since ECN-capable
    routers still drop
    packets at high levels of congestion, instead of
    continuing to send
    traffic on congested links. Thus, fot the foreseeable
    future ..."

       foreseeable future, TCP must interpret a lost
    packet as a signal of
       congestion.

       The TCP "congestion avoidance" [RFC2581] algorithm
    maintains a
       congestion window (cwnd) controlling the amount of
    data TCP may have
       in flight at any moment. Reducing cwnd reduces the
    overall bandwidth
       obtained by the connection; similarly, raising cwnd
    increases the
       performance, up to the limit of the available
    bandwidth.

    "reduces the overall bandwidth that the sender will
    attempt to use for
    the connection; similarly ..."

       TCP probes for available network bandwidth by
    setting cwnd to one
       packet and then increasing it by one packet for
    each ACK returned
       from the receiver. This is TCP's "slow start"
    mechanism. When a
       packet loss is detected (or congestion is signaled
    by other
       mechanisms), cwnd is reset to one and the slow
    start process is
       repeated until cwnd reaches one half of its
    previous setting before
       the reset.

    "This is the "slowstart threshold", or ssthresh."

       Cwnd continues to increase past this point, but at
    a much
       slower rate than before. If no further losses
    occur, cwnd will
       ultimately reach the window size advertised by the
    receiver.

     8.5 TCP Performance Characteristics

    No comments.

      8.5.1 The Formulae

    No comments.

      8.5.2 Assumptions

       Both of these formulae allow BW to become infinite
    if there is no
       loss. This is because an Internet path will drop
    packets at
       bottleneck queues if the load is too high. Thus, a
    completely
       lossless TCP/IP network can never occur (unless the
    network is being
       underutilized).

    I don't understand the last sentence in the previous
    paragraph. Is it
    saying "A completely lossless TCP/IP connection can
    never occur, unless
    the receiver is using a receive window smaller than
    the bandwidth-delay
    product"?

      8.5.3 Analysis of Link-Layer Effects on TCP
    Performance

    No comments.

    9 Quality-of-Service (QoS) considerations

    No comments.

    10 Fairness vs Performance

    No comments.

    11 Delay Characteristics

    No comments.

    12 Bandwidth Asymmetries

       Some subnetworks may provide asymmetric bandwidth
    (or may cause TCP
       packet flows to experience asymmetry in the
    capacity) and the
       Internet protocol suite will generally still work
    fine. However,
       there is a case when such a scenario reduces TCP
    performance. Since
       TCP data segments are 'clocked' out by returning
    acknowledgments TCP
       senders are limited by the rate at which ACKs can
    be returned
       [BPK98]. Therefore, when the ratio of the
    bandwidth of the
       subnetwork carrying the data to the bandwidth of
    the subnetwork
       carrying the acknowledgments is too large, the slow
    return of of the
       ACKs directly impacts performance. Since ACKs are
    generally smaller
       than data segments, TCP can tolerate some
    asymmetry, but as a general

    Should we say "TCPs can tolerate substantial
    assymetry, ..."?

       rule designers of subnetworks should be aware that
    subnetworks with
       significant asymmetry can result in reduced
    performance, unless
       issues are taken to mitigate this [asym-id].

       Several strategies have been identified for
    reducing the impact of
       asymmetry of the network path between two TCP end
    hosts, e.g. [asym-
       id]. These techniques attempt to reduce the number
    of ACKs
       transmitted over the return path (low bandwidth
    channel) by changes
       at the end host(s), and/or by modification of
    subnetwork packet
       forwarding. While these solutions may mitigate the
    performance issues
       caused by asymmetric subnetworks, they do have
    associated cost and
       may have other implications. A fuller discussion of
    strategies and
       their implications is provided in [asym-id].

    The IESG documents under review is showing Allison
    holding this token.

    13 Buffering, flow & congestion control

    No comments.

    14 Compression

       We make a stronger recommendation that subnetworks
    operating at low
       speed or with small MTUs compress IP and
    transport-level headers (TCP
       and UDP) using several header compression schemes
    developed within
       the IETF. An uncompressed 40-byte TCP/IP header
    takes about 33
       milliseconds to send at 9600 bps. "VJ" TCP/IP
    header compression

    Can we introduce this as "Van Jacobson TCP/IP header
    compression", to be
    consistent with what we settled on in 2.5G/3G?

       [RFC1144] compresses most headers to 3-5 bytes,
    reducing transmission
       time to several milliseconds. This is especially
    beneficial for
       small, latency-sensitive packets in interactive
    sessions.

    15 Packet Reordering

    No comments.

    16 Mobility

    No comments.

    17 Routing

    No comments.

    18 Security Considerations

    Too many transitions in this paragraph?

       Nonetheless, end-to-end security mechanisms are not
    used as widely as
       might be desired. However, the group could not
    reach consensus on
       whether subnetwork designers should be actively
    encouraged to
       implement mechanisms to protect user data.

    I propose:

       Although end-to-end security mechanisms are not
    used as widely as
       might be desired, the PILC working group could not
    reach consensus on
       whether subnetwork designers should be actively
    encouraged to
       implement mechanisms to protect user data.

    And can we drop the "however, " in this paragraph? It
    doesn't seem to be
    in contrast to anything...

       However, traffic analysis is a notoriously subtle
    and difficult
       threat to understand and defeat, far more so than
    threats to
       confidentiality and integrity. We therefore urge
    extreme care in the
       design of subnetwork security mechanisms
    specifically intended to
       thwart traffic analysis.

    Not to be pedantic, but [WFBA2000] isn't about buffer
    overruns in I/O
    system calls - it's about buffer overruns in the
    standard C library:

       In addition, well-designed security protocols can
    be compromised by
       implementation defects. Examples of such defects
    include use of
       predictable pseudo-random numbers [RFC1750],
    vulnerability to buffer
       overflow attacks due to unsafe use of certain I/O
    system calls
       [WFBA2000], and inadvertent exposure of secret
    data.

    References

       [AP99] M. Allman, V. Paxson, On Estimating
    End-to-End Network Path
       Properties, In Proceedings of ACM SIGCOMM 99.

    A URL for this paper is:
    http://www.acm.org/sigcomm/sigcomm99/papers/session7-3.html

       [AR02] G. A,car and C. Rosenberg, Weighted Fair
    Bandwidth-on-Demand
       (WFBoD) for Geo-Stationary Satellite Networks with
    On-Board
       Processing, Special Issue on Broadband Satellite
    Systems: A Network
       Perspective, Computer Networks, accepted on July
    13, 2001. To appear
       in 2002.

    The best URL I could find for this paper was
    http://www.iis.ee.ic.ac.uk/~g.acar/Research.html#WFBoD

       [asym-id] H. Balakrishnan, V. N. Padmanabhan, G.
    Fairhurst, M,
       Sooriyabandara. "TCP Performance Implications of
    Network Path
       Asymmetry", work in progress as internet-draft,
    draft-ietf-pilc-
       asym-07.txt, November 2001.

    This draft should have already expired. Aaron - are we
    going to progress
    it?

       [ARQ-DRAFT] Fairhurst, G., and L. Wood, Advice to
    link designers on
       link Automatic Repeat reQuest (ARQ), work in
    progress as internet-
       draft, draft-ietf-pilc-link-arq-issues-03.txt,
    August 2001.

    This is now version 04, from March 2002.

       [ATMFTM] The ATM Forum, "Traffic Management
    Specification, Version
       4.0", April 1996, document af-tm-0056.000
    (www.atmforum.com).

    A more complete URL is
    ftp://ftp.atmforum.com/pub/approved-specs/af-tm-0056.000.pdf

       [BA02] Ethan Blanton, Mark Allman. On Making TCP
    More Robust to
       Packet Reordering. ACM Computer Communication
    Review, 32(1), January
       2002.

    A complete URL is
    http://www.acm.org/sigcomm/ccr/archive/2002/jan02/ccr-200201-allman.pdf

       [BPK98] Hari Balakrishnan, Venkata Padmanabhan,
    Randy H. Katz. 'The
       Effects of Asymmetry on TCP Performance." ACM
    Mobile Networks and
       Applications (MONET), 1998.

    A complete URL is
    http://daedalus.cs.berkeley.edu/publications/tcpasym-mobicom97.ps.gz

       [Crypto9912] Schneier, Bruce "European Cellular
    Encryption
       Algorithms" Crypto-Gram (December 15, 1999)
    www.counterpane.com

    A complete URL is
    http://www.counterpane.com/crypto-gram-9912.html

       [DOCSIS3] W.S. Lai, "DOCSIS-Based Cable Networks:
    Impact of Large
       Data Packets on Upstream Capacity", 14th ITC
    Specialists Seminar on
       Access Networks and Systems, Barcelona, Spain,
    April 25-27, 2001.

    A complete URL is
    http://www.att.com/networkandperformance/docs/ndrp_itc0401b.doc

       [ES00] David A. Eckhardt and Peter Steenkiste,
    "Effort-limited Fair
       (ELF) Scheduling for Wireless Networks, Proceedings
    of IEE Infocom
       2000.

    The closing quote mark is missing (after "Networks").
    A complete URL is
    http://www.ieee-infocom.org/2000/papers/266.ps

       [FB00] Firoiu V., and Borden M., "A Study of Active
    Queue Management
       for Congestion Control" to appear in Infocom 2000

    The paper has already appeared. The URL is
    http://www.ieee-infocom.org/2000/papers/405.pdf

       [IEEE80211] IEEE 802.11 Wireless LAN standard.
    Available from
       http://standards.ieee.org/catalog/IEEE802.11.html.

    This is now freely downloadable from
    http://standards.ieee.org/getieee802/download/802.11-1999.pdf

       [LKJK02] R. Ludwig, A. Konrad, A. D. Joseph, R. H.
    Katz, "Optimizing
       the End-to-End Performance of Reliable Flows over
    Wireless Links",
       Kluwer/ACM Wireless Networks Journal, Vol. 8, Nos.
    2/3, pp. 289-299,
       March-May 2002.

    A Complete URL is
    http://iceberg.cs.berkeley.edu/papers/Ludwig-Mobicom99/

       [LRKOJ99] R. Ludwig, B. Rathonyi, A. Konrad, K.
    Oden, A. Joseph,
       Multi-Layer Tracing of TCP over a Reliable Wireless
    Link, pp.
       144-154, In Proceedings of ACM SIGMETRICS 99.

    Need quotes around title. A complete URL is
    http://iceberg.cs.berkeley.edu/papers/Ludwig-Sigmetrics99/

       [LS00] R. Ludwig, K. Sklower, The Eifel
    Retransmission Timer, ACM
       Computer Communication Review, Vol. 30, No. 3, July
    2000.

    Need quotes around title. A complete URL is
    http://www.acm.org/sigcomm/ccr/archive/2000/july00/ccr_200007-ludwig.html

       [MBB00] May, M., Bonald, T., and Bolot, J-C.,
    "Analytic Evaluation of
       RED Performance" to appear INFOCOM 2000

    The paper has appeared. A complete URL is
    http://www.ieee-infocom.org/2000/papers/369.ps

       [MBDL99] May, M., Bolot, J., Diot, C., and Lyles,
    B., "Reasons not to
       deploy RED", technical report, June 1999.

    A complete URL is
    http://www-sop.inria.fr/rodeo/personnel/mmay/may_red.html

       [GM02] Luigi Alfredo Grieco1, Saverio Mascolo, "TCP
    Westwood and Easy
       RED to Improve Fairness in High-Speed Networks",
    Proceedings of the
       7th International Workshop on Protocols for
    High-Speed Networks,
       April 2002.

    A complete URL is
    http://www.cs.ucla.edu/NRL/hpi/papers/2002-pfhsn-0.pdf.gz

       [MSMO97] M. Mathis, J. Semke, J. Mahdavi, T. Ott,
    "The Macroscopic
       Behavior of the TCP Congestion Avoidance
    Algorithm", Computer
       Communication Review, volume 27, number 3, July
    1997.

    A complete URL is
    http://www.psc.edu/networking/papers/model_abstract.html

       [PFTK98] Padhye, J., Firoiu, V., Towsley, D., and
    Kurose, J.,
       "Modeling TCP Throughput: a Simple Model and its
    Empirical
       Validation", UMASS CMPSCI Tech Report TR98-008,
    Feb. 1998.

    A complete URL is
    ftp://gaia.cs.umass.edu/pub/Padhye-Firoiu98-TCP-throughput-TR.ps

       [RF95] Romanow, A., and Floyd, S., "Dynamics of TCP
    Traffic over ATM
       Networks". IEEE JSAC, V. 13 N. 4, May 1995, p.
    633-641.

    A complete URL is
    http://www-nrg.ee.lbl.gov/papers/tcp_atm.pdf

       [RFC1812]

    Needs details!

       [Schneier3] Schneier, Bruce "Why Cryptography is
    Harder Than it
       Looks", www.counterpane.com

    A complete URL is
    http://www.counterpane.com/whycrypto.html

       [TCPF98] Dong Lin and H.T. Kung, "TCP Fast Recovery
    Strategies:
       Analysis and Improvements", IEEE Infocom, March
    1998. Available
       from:
    "http://www.eecs.harvard.edu/networking/papers/infocom-tcp-
       final-198.pdf"

    This URL is broken:

       [WFBA2000] David Wagner, Jeffrey S. Foster, Eric
    Brewer and Alexander
       Aiken, "A First Step Toward Automated Detection of
    Buffer Overrun
       Vulnerabilities", Proceedings of NDSS2000, or
       www.berkeley.edu:80/~daw/papers/

    A correct URL is http://www.cs.berkeley.edu/~daw/papers/overruns-ndss00.ps

    __________________________________________________
    Do You Yahoo!?
    Yahoo! - Official partner of 2002 FIFA World Cup
    http://fifaworldcup.yahoo.com

    _______________________________________________
    pilc mailing list
    pilc@ietf.org
    https://www1.ietf.org/mailman/listinfo/pilc
    http://www.ietf.org/html.charters/pilc-charter.html
    http://pilc.grc.nasa.gov/



    This archive was generated by hypermail 2b29 : Tue Jun 11 2002 - 13:43:30 EDT