William> On Tue, 7 Jan 1997, Patrick Schaaf wrote:
[snip]
William> In a quick summary: one might take an order or two less time checking the
William> page than reading from or writing to disk, depending on the disk and CPU.
William> Someone suggested to me that a system running X, httpd and few users might
William> swap about 6-10% zero filled pages. This is of more interest to users
William> doing calculations with large sparse matrices or other computations with
William> large zero filled buffers than home users.
<RANT>
On a releated topic (Large Matrices) the biggest way to improve large
matrix calculations. Especailly multiplication, is to use a blocking
technique so their locality of referrence improves. This should
improve for some people at least, their paging performance
emmenciely. Heck it should even improve the L1 cache performace that
it was designed to improve.
There are other improvements to take advantage of the sparseness of
matrices if you know it is going to occur.
Sorry I have seen one to many rant about my matrix calculation pages
hard on linux-kernel....
</RANT>
Actually with checking to see if a page is zero filled the proper
place to do it would be, when it is occasionally made inaccessible but
still in the system to get usage information. That way even high
usage zero pages could benefit from this optization. The question is
of course how much of a speed penalty would it entail to it it there?
William> I am poking at relevent kernel files to see what is going on, but any
William> pointers would be strongly appreciated.
Sorry None from me.
Eric