>
> Around slide # 13, Van Jacobson said that the 4K initial window was a gre=
> at
> idea. He also said that counting bytes rather than ACKs was a bad idea a=
> nd
> would produce bursts. There is a need to spread timing at the sender.
> Ssthresh tells you the buffer limit in the bottleneck. Need the right
> solution to an intermediate small buffer.
In certain situations (e.g., network with a high degree of bandwidth
asymmetry, such as in certain satellite-based systems), it is desirable to
decrease the frequency of acks (i.e., have one ack sent for every n data
segments, where n could be larger than 2). Indeed, even if we don't decrease
the ack frequency deliberately, this could happen simply because the
bottleneck buffer along the ack path fills up and drops several acks.
However, since this is NOT a reflection of the state of the data path,
I believe it is a reasonable goal to ensure that the TCP sender's window
growth is not adversely affected. This requires that the sender also
take into account the acks that got dropped, i.e. count bytes.
It is certainly the case that the sender should avoid bursting out data,
presumably by using timers. A simple way of computing the timer is to
start with the "rate" of the connection, cwnd/srtt, and then compute
how long to wait before sending out the next chunk of data. To work
well, this technique requires a fairly accurate estimate of srtt, which
can be obtained in practice by appropriately using the timestamp option.
This technique is discussed in a paper Mobicom '97 paper that I have
co-authored with Hari Balakrishnan and Randy Katz. The paper is available
from:
ftp://daedalus.cs.berkeley.edu/pub/papers/tcpasym-mobicom97.ps (914 KB)
ftp://daedalus.cs.berkeley.edu/pub/papers/tcpasym-mobicom97.ps.gz (192 KB)
The results reported in this paper were obtained from simulation in
the ns simulator. We are currently in the process of implementing
many of the algorithms in the BSD/OS 3.0 kernel.
> Van Jacobson discussed results from 1986 or 87 from SATNET that are writt=
> en
> down in some unspecified seminar notes. He suggested that an ACK spacing
> box at a downlink ground station could be used to clock the sender to avo=
> id
> bursts. Tim Shepard pointed out that if a sender already has packet
> spacing it should turn it off if there is an intermediate ACK spacer box =
> in
> the link. Van replied that his box would only space ACKs when it saw bur=
> sts.
Relevant to this, a company called Packeteer (www.packeteer.com) sells
a box called PacketShaper that shapes TCP traffic (to enforce bandwidth
allocation) by manipulating the spacing between acks.
> Peter Warren claimed that the asymmetry of HTTP data is on the order of
> 6:1. He said this is a potentially serious bottleneck on the upstream
> link, would like to hear form others working on asymmetric links. The
> meeting was then adjourned.
>
Bidirectional traffic over asymmetric networks brings forth several
performance problems. Some of these are analysed in our Mobicom '97
mentioned above.
-Venkat
This archive was generated by hypermail 2b29 : Mon Feb 14 2000 - 16:14:30 EST