> With a smaller stdio buffer, you do more read/write syscalls and add more > overhead. Ah yes, I see, thanks.
> The [time spent] in a syscall is attributed to the process. But the > actual I/O is often deferred to kernel threads and handled in > interrupts. Interrupt time is not accounted to user processes. While I do know this in principle, I'm not familar enough with callouts, wokqueues, kernel threads etc. to tell where in my case (FFS on RAIDframe on mfii) time is attributed to the process and where to the kernel. For instance, when reconstructing (I manually failed a component to test this), top shows the "system" process to use 80-90% of CPU time, so I guess the EXORing done during reconstruction is attributed to the kernel. But during normal writes, who is attributed the time to compute the parity?
