Comment 32 for bug 709245

I tried to gather more data on this and made several runs of "hdparm -Tt" for the good and bad case:
- For the good case (nosmp)
  - cached reads avg. 184.61MB/sec (MIN=181.02 MAX=186.30 STDDEV=1.62)
  - buffered reads avg. 3.64MB/sec (MIN=3.64 MAX=3.65 STDDEV=0)
- For the bad case
  - cached reads avg. 62.00MB/sec (MIN=21.98 MAX=111.55 STDDEV=27.97)
  - buffered reads avg. 1.35MB/sec (MIN=1.09 MAX=1.86 STDDEV=0.20)

So not only the SMP case is slower, there is also a high variation in the numbers. Whatever happens, it is not a linear slowdown.

When trying various options maxcpus=1 has the same effect as nosmp (disabling a core later does not improve things), there was one option (which right now slips my memory) causing a higher rate of timer interrupts which lead to a even worse buffered reads performance (as was disabling the irqbalance daemon). Using nohlt seemed to have no effect.

When looking at the interrupts, the main difference seemed to be the use of gp_timer when booted with nosmp or maxcpus=1 and LOCal timers in the other case. Booting with 2 CPUs and disabling one did not change that. I cannot recall whether the IPIs stopped being incremented completely or only on the other CPU. But LOC definitely was still used for CPU#0.