The general question for memory efficiency is how many bytes per user are required
in an ICEfaces application vs a stock JSF application. How
much memory is used by:

 JSP tags
 Facelet handlers
 Component tree
 DOM
 Auxiliary objects (View, servlets, ...)

Problem areas:

The most significant problem we're already aware of is with
JSF state saving. ICEfaces keeps the component tree in memory
for every user. Theoretically this reduces latency (which we
should measure) but the downside is that more memory is require
for the component trees than the condensed representation as
arrays of name/value pairs. This difference also introduces
behavioral differences and problems with failover.

A second significant difference (that likely will not be revealed
by a CPU profiler because it doesn't change the distribution)
is that ICEfaces component trees are significantly larger due
to the use of a component for each XML tag in the input document.
How much larger is what we need to measure. This increase in
component tree size yields greater CPU and memory consumption at
every stage, so it's important to tackle if we have the time
(here's where we want to use the SAX event reply rather than
individual components for static text).