Degree of memory sharing -- anecdotal data.
Jonathan S. Shapiro
Sun, 8 Feb 1998 14:01:41 -0500
Norm has argued strenuously for the importance of shared mappings.
Here is a solitary bit of supporting evidence.
I have just learned about a tool called xosview that runs on Linux.
This tool displays various system statistics, including CPU, swap
space, and memory utilization. More importantly, it also shows the
usage breakdown of these statistics.
My build machine at home is a 32 Mb, 120Mhz Pentium.
While rebuilding the EROS tree, memory utilization total is 97%+ once
things get rolling. CPU hovers around 90% -- mostly delayed for file
access. Swap area usage hovers at 31%, but the idle usage is 28%.
Now for the interesting bit -- the portion of memory that is reported
as "shared" hovers around 65% during the build, where the normal level
is closer to 40%. Shared memory includes dynamic library code and
unmodified data, code segments (e.g. for multiple shells) and the
UNIX, mind you, promotes sharing because of the multiplicity of shells
and dynamic libraries. Nonetheless I found this number staggering.
Indirectly, it supports the contention that shared mapping tables for
these structures are important to system performance.