CURRENT MEETING REPORT
Reported by Bob Braden, USC Information Sciences Institute
Minutes of the Transaction TCP BOF (ttcp)
This BOF was convened to discuss a set of TCP extensions for efficient
transaction-mode communication, i.e., request-response mode, which
is currently documented in an Experimental RFC [Braden, R., "T/TCP
-- TCP Extensions for Transactions, Functional Specification",
RFC-1644, ISI, July 1994]. There has recently been interest in
using T/TCP, especially as a component of performance improvements
for the World Wide Web. As a result, some have suggested starting
a working group to create a standard.
This BOF was convened to discuss T/TCP and its possible standardization.
The main speaker was Bob Braden, who developed T/TCP as an NSF-funded
research project. This talk covered the history of early efforts
to define an Internet standard transaction transport protocol,
the development of TCP, a brief outline of the theory and features
of T/TCP, and finally its advantages and disadvantages. See the
slides following.
A number of interesting issues were raised.
T/TCP optimizes transactions after the first between any pair
of hosts. However, some transaction-mode applications, e.g., the
DNS, have the property that a given client makes individual transactions
to a large number of different servers, in which case T/TCP will
not provide much performance improvement over pure TCP.
To provide efficient transaction transport in that case, synchronized
clocks are needed. It was suggested that one might consider synchronizing
clocks for DNS server hosts, a small subset of all Internet hosts.
More work may be needed on T/TCP congestion control. As a a compatible
extension of TCP, T/TCP includes all the normal TCP congestion
control machinery, which comes into play as transactions become
larger. However, it was observed that TCP has no congestion control
on SYN segments, and if there is heavy use of T/TCP for minimal
transactions, the Internet will be filled with SYN segments. It
may be necessary to include some rate-based control over transactions,
in addition to the normal byte-based congestion control of TCP.
One view of T/TCP is that each request and each response is a
logical "record". In a stream of successive transactions
between the same (client,server) pair, connection termination
is effectivley being used as a record mark. An alternative way
to get the record mark functionality would be to insert a "thin"
protocol layer above TCP, a "session" layer", whose
framing carries a record mark. This alternative is being explored
for the World Wide Web.
Christian Huitema recently proposed a TCP modification in which
each packet carries logical connection handle, to better support
mobility and dynamic host configuration. Since T/TCP also introduces
a logical connection handle, an incarnation number or "connection
count", it may be possible to use the same handle for both.
The group consensus was that it would be useful to have a working group to consider standardization of T/TCP. Roughly a dozen people out of the 60 attendees indicated that they would participate in such a working group. The chair promised to report this to the Area Director.