| From: Robert P. J. Day <rpjday@xxxxxxxxxxxxxx> | maybe someone more knowledgeable can clarify this. and, again, i'm | still unsure if there's any harm in running a PAE-enabled kernel even | when you don't need one. There would only be one kernel if supporting PAE didn't have a cost. There is probably a cost in code size and in code pathlength. I found this datapoint: http://linux-memhotadd.sourceforge.net/hotadd.html "Benchmarks seem to indicate around 3-6% CPU hit just for using the PAE extensions (ie. it applies regardless of whether you are actually accessing memory locations greater then 4GB)." I don't know if it is reliable -- read the paper and you judge. I would guess that the real CPU time cost is some linear combination of the number of system calls, interrupts, and process switches. For example, a system with only CPU-intensive applications might not be hit. Now that I read more of the paper, that seems to be confirmed -- see 6.1.1.
