It seems Federico enthused a chap called Ashley
Pittman with my plan to create an I/O-grind type implementation:
think cachegrind but for I/O, like cachegrind there are two parts:
Sampling
- First separate all time into 1/2 seek time
intervals (cf. H. Nyqist), ie. ~5ms intervals.
- in each interval track & log program behavior:
- all explicit I/O syscalls: eg.
open/read/write/stat etc.
- track all implicit I/O: 1st time writes to
new pages, touches of mmapped files (code & data).
- generate a log-file with filename / offset / size &
stack-trace entries (elide lots of call traces
across this interval).
Processing / visualisation
Anyhow - the next part is turning this data into meaningful
performance numbers, in a deterministic fashion - ie.
we need to be able to run the same data-set twice and get the
same number:
- a standard disk I/O latency simulator apparently
there are a lot of these around the place.
- simple (deterministic, but statistical?) filesystem
model (to turn paths to block numbers).
- noddy swap / LRU page replacement algorithm.
- simple page-cache / inode / dentry cache simulation.
- mindless [ and deterministic ] process scheduler.
Impact
So - of course the simulations don't have to be particularly
advanced, or faithful to the O/S to give interesting results
I think: much in the way that callgrind's 'cycle estimation'
is a helpful guesstimate. That's important, as while the valgrind
piece is hard to write, creating super-accurate simulators of
kernel behavior is rather harder.
It is my contention though, that cold-start (and hence login
time, boot time etc.) remain such untractible problems due to a
lack of profiling tools for unexpected I/O behavior:
particularly excessive seeking, and/or just reading too much
data. Without tools to measure this - and particularly to be
able to repeatably (and ~instantly) profile different data sets,
with different (kernel) algorithms would make a huge difference.
Though it is indeed a KDE program, KCachegrind rocks my
world, integrating with that beasty, and allowing the profiling
to be tweaked eg. having a drop-down selection of: "aggregate
I/O latency", "explicit I/O latency", "page dirties", "I/O bandwidth"
etc. would be wonderful, also some model parameters to tweak:
a "system memory" spin-button, a coarse disk characteristic:
"Laptop" vs. "Desktop", a filesytem button: "ext3" vs.
"ReiserFS" etc. etc.
