It was a fun meeting. The Apple campus is a great place and they
have a late night bar nearby so we got some food afterwords.
I have put the slides up on the DragonFly site:
http://www.dragonflybsd.org/docs/
http://www.dragonflybsd.org/docs/LISA200512/
They are fairly self explanatory. There are a couple of things to
note:
* Keep in mind that the TSC's on the two cpus drift from each other. I
was resynchronizing them 10 times a second but they are still upwards
of 200ns off. This is demonstrated in the IPIQ Messaging PING-PONG
test numbers.
* Timings that show repetition with sequential sequence numbers (on any
given cpu) tend to be end-to-end tests. So, e.g. the tsleep/wakeup
tests are end-to-end tests that include scheduler overheads.
* There are two sets of SLAB allocator free() path tests. The first
set has bogus numbers (its on a slide so I could explain why they
were bogus, which was simply due to the extensive KTR logging blowing
up the numbers).
I was quite impressed with the results. I did not expect the IPI
code to be so fast, and even medium-complexity code such as processing
a TCP packet tended to take on the order of ~1 uS. That's a good
argument for not doing preemption of non-interrupt threads.
The major point of the presentation was to show the efficiencies gained
when operations can be aggregated. e.g. the use of passive IPI message
queueing for free() tends to aggregate several free operations on the
target cpu which are then executed in a tight loop. Same with the network
RX interrupt and both network and TCP packet processing. Etc.
-Matt
Matthew Dillon
<dillon at xxxxxxxxxxxxx>