next up previous contents
Next: Comparison with another RSVP Up: Evaluation Previous: Performance   Contents


Comparison with standard ATM signalling

Tests [#!DataCom-ATM-tests!#] about ATM switches were recently made by Data Communications magazine. One of the tests consisted in measuring the number of connections setup and then torn down by the switches. Early signalling software spent about 100ms per call. Today's modern switches make at least 100 calls per second; Fore's ASX-1000 was reported to set up and tear down around 190 calls per second.

The RSVP process can process a Path message, a Resv message and a Path Teardown message in $\delta_{RSVP}$ (20ms, say).

An Ariel Setup/Teardown operation ( $\delta_{Ariel}$) has been measured [#!OSSA!#] as less than 9ms (this delay depends on the way the RPC4.3 is done); however, we need to do two different RPCs: the first one is done in order to establish the connection during the processing of the Path message, and the other one is done later in order to tear down the connection when a Teardown message is received.

Several network accesses need to be done: for each message, the packet needs to reach the switch ( $\delta_{src \rightarrow sw}$), then it has to be forwarded to the control architecture ( $\delta_{sw \rightarrow ca}$), then the control architecture sends a RSVP message back to the switch ( $\delta_{ca \rightarrow sw}$), and the switch sends it to the receiver ( $\delta_{sw \rightarrow dst}$). This means that we have four network accesses for each RSVP message. A network access is estimated as 1ms.

The time spent to establish and then tear down a connection is therefore:


\begin{displaymath}\displaystyle\Delta =
\sum_{\parbox{1cm}{\centering \scriptsi...
...size Time spent to setup and then teardown an ATM connection}}
\end{displaymath}


\begin{displaymath}
\simeq 3 * 4ms\;+\;20ms\;+\;2 * 9ms = 50ms
\end{displaymath}

The control architecture designed can perform about \(\frac{1}{\Delta} \simeq 20\) calls per second. Doing ATM signalling in this way gives us open control of the switch as well as RSVP capabilities, so these results are fair. This control architecture is between 2 and 3 times slower4.4 than the Hollowman [#!Hollowman!#], another open ATM control architecture, and it is about 10 times slower than today's most efficient switches.


next up previous contents
Next: Comparison with another RSVP Up: Evaluation Previous: Performance   Contents
Hugo Haas
hugo@larve.net
July 1998 - Please note that this HTML version is broken; I advise you to read the PostScript version.