Cell Relay Archive

[Srikrishna <m_srikrishna@fujitsu-fnc.com>]

William Stallings wrote:

> In article <6csnbt$p2q$1@nnrp2.dejanews.com>, rhiggins@westell.com wrote:
>
> > Hello:
> >
> > Can somebody explain to me what PPD is?  And how it is different from EPD.
> >
> > Thanks In Advance
> >
> > Rodger Higgins
> > rhiggins@westell.com
> >
>
> These terms refer to stratgies for carrying TCP/IP over ATM. Here is an
> excerpt from my book that discusses them. (HIGH-SPEED NETWORKS: TCP/IP and
> ATM Design Principles,Prentice-Hall, William Stallings, 1998; ISBN
> 0-13-525965-7). Sorry, I can't include the figures.
>
> ------------------------------------------------------------------------
>    Partial Packet Discard and Early Packet Discard
>    The conclusions of all of these studies have led a number of switch
> vendors to increase ATM switch buffer size. A number of vendors have also
> been influenced by proposals for two techniques to improve the cell
> discard behavior. These proposals are referred to as partial packet
> discard and early packet discard. Both strategies are aimed at reducing
> the transmission of useless cells.
>    Partial packet discard (PPD) operates as follows: if a cell from an IP
> datagram is dropped from a switch buffer, the subsequent cells in that
> datagram are also discarded. For a switch to recognize, at the ATM level,
> which cells belong to a given IP datagram, PPD must operate on a
> per-virtual-channel basis. When a switch discards a cell on a particular
> VC, it then discards all subsequent cells on that VC until it sees a cell
> with the SDU type bit set to one in the ATM cell header. This cell marks
> the end of an AAL5 PDU and therefore the end of an IP datagram. This final
> cell is not dropped. Because AAL5 does not support the multiplexing of
> cells from different PDUs, the SDU parameter can be used successfully to
> delimit IP datagram boundaries.
>    Figure 10.16c shows the performance of PPD. As can be seen, PPD
> provides better performance than TCP over plain ATM, but is not as
> effective as one might wish. Improvement is limited because PPD only
> discards the "tail end" of a datagram. On average, we can expect that only
> half of a damaged datagram is discarded with each cell discard.
>    A more effective scheme is early packet discard (EPD). In this case,
> when a switch buffer reaches a threshold level, but before it is actually
> required to discard any cells due to buffer overflow, an entire IP
> datagram is dropped. Thus, when the switch senses that congestion is
> beginning and that cell discard may soon be necessary, it preemptively
> discards all of the cells, starting with the first, of an IP datagram. For
> this purpose, the switch looks for the first incoming cell on a VC with an
> SDU bit of 0 that immediately follows a cell on the same VC with an SDU
> bit of 1. This cell marks the beginning of a new datagram, and the
> discarding begins with this cell. In effect, the EPD strategy emulates a
> packet-switching network in which entire packets are dropped.
>    Figure 10.16d shows the performance of EPD, with a threshold set to
> half the total buffer size. Except for very small buffer sizes, high
> effective throughput is achieved.
>    The combination of EPD with the TCP congestion control mechanism seems
> to provide effective throughput for the average TCP connection. EPD has a
> short-term orientation, anticipating congestion and reacting immediately
> to drop a packet. Complementing this action is the longer-term orientation
> of TCP, which reacts to packet loss by backing off and only slowly
> resuming its former rate of transmission.
> ------------------------------------------------------------------------
>
> |                | Descriptions, errata sheets and discount order info |
> | Bill Stallings |              for my current books and               |
> |  ws@shore.net  |       information on my forthcoming books at        |
> |                |              http://www.shore.net/~ws               |

Justwant to clarify a subtle point to the above explanation.

If the first cell of an IP packet is discarded using the PPD mechanism then I
would assume that this is same as EPD because all the cells corresponding to
that IP packet are going to be discarded before they are queued. However there
is no guarantee that this will be the case. Most probably, some intermediate
cell of an IP packet may get discarded due to congestion with earlier cells of
this IP packet already queued in the switch. As a result, junk cells are going
to be switched unnecessarily  because TCP is going to retransmit the entire IP
packet again anyway. Hence PPD is less efficient than EPD.

EPD on the other hand checks to make sure that the entire IP packet (and hence
all the cells pertaining to this IP packet) has enough buffer resources upfront
(via athreshold mechanism) before queueing any cell for this IP packet. If so,
then no cell (starting one or intermediate ones) are discarded as a result of
which junks cells as mentioned above are not switched. Hence EPD is more
efficient than PPD.

However, both PPD and EPD require VC queueing implementation in the switch I
presume?

Having said that EPD is more efficient than PPD, note that it is only within
that switch. If a subsequent switch drops the entire packet, the endpoint TCP
will retransmit the entire IP packet which makes the first switch look
inefficient but not the subsequent switch whihc caused the retransmit.

Any comments or corrections to this is greatly appreciated to ensure what I
state above is not totally wrong.

Thanks in advance..

Srikrishna..