virtual memory 4511
That number came from zooming in on some lines on graph in a fuzzy copy of a document. There are multiple lines on the graph, covering multiple scenarios, and other parts of the lines are further apart.
The point is that at some points it get so close as to be hard to differentiate from LRU. More below.
virtual memory 4512
fOn Mon, 15 May 2006 22:38:39 +0200 in alt.folklore.computers, Terje Pool-subpool allocation seems advantageous in situations where...
That was my example to get the point across as to why it works. Yes, no one would have a 1 page working set.
The free list (the list of pages evicted from the WS) tends to LRU ordering because more recent pages get pulled out of it and back into the working set list. When a frame must be rebuttigned to satisfy a hard fault to disk, the head of the free list will be the LRU page.
There is also a competing cost because the number of soft faults rises as the working set size shrinks.
In my extreme example of a 1 page WS, every memory reference causes a fault, which evicts the frame from the WS to the tail of the free list and reclaims a valid frame from someplace in the free list. The free list winds up in perfect LRU order. This also has the highest cost.
The paper analyzed two programs: a Fortran compilation consisting of 4 million references to 359 (512-byte) pages, and a Bliss compilation consisting of 8.5 million references to 510 pages.
It measured the fault rates and costs for segmenting different amounts (10%, 20%, 30%) of the memory off onto the free list for various amounts of available memory.
Zooming in on the graph for a 30% free list split with 150 pages shows the fault count line so close it touches the LRU line.