> 	- Too much dirty data causes saturation of the
> 	  I/O devices

If you flush it all at once...

From Nancy Cluts' description of the design principles behind the NT filesystem
architecture I seem to remember the "sync" thread runs about every four seconds,
but only flushes one quarter of the dirty pages at one run (this thread does 
other things like file compression too, so the real implementation must be 
more complicated than this simple description). This should avoid longer I/O 
saturation and make the flushing invisble to the user.

Their existing implementation does not fulfill this theoretical promises, I've
seen my NT workstation freeze for periods of about 45 seconds with heavy
disk activity and no foreground jobs doing any real work.


Martin