Greater resolution mem_load from contest
Here are some results you may find informative.
I adjusted the mem_load module of contest (http://contest.kolivas.net) to vary the memory load from 10 to 110% in increments of 10% and performed the test on the following kernels.
What I found is that performance regardless of the kernel was constant (for that kernel) up to 70% (presumably the critical number for my 256Mb machine).
What happened beyond this point, however, was quite different between kernels.
Here are the results:
Kernel: 2.4.19 rmap14b 2.5.36 2.5.36-mm1
Mem%
60 76.34 76.98 66.62 69.12
70 76.22 77.14 67.21 67.28
80 79.20 80.14 68.24 70.29
90 82.59 115.51 148.63 92.96
100 84.21 108.61 107.54 95.50
110 92.49 114.51 132.45 95.32
As you can see, in absoute performance the 2.5 kernels at low mem loads are better.
rmap14b (2.4.19-rmap14b) performance is identical to vanilla till 80%. Beyond this point it starts to deteriorate rapidly.
2.5.36 exhibits this same behaviour (presumably for the same reason?). Note the dip at exactly 100% and the peak either side of it? -mm1 seems to do better than vanilla 2.5.36
Overall, vanilla 2.4.19 seems to respond more graded.
A quick reminder what these numbers are;
the data value is the time taken to compile a kernel, and the mem% is a background memory load that continually asks for x% of the memory.
I'd like to include the ability to test this into a newer version of contest; however the critical point when the results start to deteriorate, and the absolute resolution required to show the difference will be dependent on the test machine's memory. I haven't resolved the best way around this.
Con.