usertest notes

[Martin Pool]

On Fri, Sep 26, 2008 at 1:57 PM, Robert Collins
<robertc at robertcollins.net> wrote:
> I'd love to see networking ones to aid the stuff andrew is working on.

Me too; let me sketch out the path to that.

We need to add some test scenarios (according to
http://people.ubuntu.com/~ianc/plugins/usertest/doc/usertest-tutorial.html)
for network operations.

We also need to choose whether to measure them in real elapsed time or
by counting roundtrips.  For things inside selftest I think counting
effort is best; for usertest measuring real time is good.  To use the
same method as in <http://bazaar-vcs.org/SmartPushAnalysis1.4> we can
use netem, which must be configured by root, so we need to ask the
sysadmins to set it up on escudero.   And of course we have to
actually set up

<http://www.linuxfoundation.org/en/Net:Netem> shows how we can filter
based on destination address with something like this:

  tc qdisc add dev lo root handle 1: prio
  tc qdisc add dev lo parent 1:3 handle 30: netem delay 500ms
  tc qdisc add dev lo parent 30:1 handle 60: tbf rate 100kbit buffer
1600 limit 3000
  tc filter add dev lo protocol ip parent 1:0 prio 3 u32 match ip
dport 4155 0xffff flowid 60:1

This is not precisely the same as what we were previously using
because it has a bandwidth limit as well as a delay limit.  (In
practice I believe the previous delay limit was such that tcp
windowing stopped it every using the full nearly-infinite bandwidth.)

I'm not sure that having the filter to port 4155 is really needed, as
this machine does nothing else but run these benchmarks and nothing
should be harmed by having the slowest loopback device in the world.
Possibly leaving it out is more flexible.

Then it would be enough to have

# tc qdisc add dev lo root handle 1: netem delay 500ms
# tc qdisc add dev lo parent 1: handle 30: tbf rate 100kbit buffer
1600 limit 3000

On consideration I'll ask them to give us a command to turn this on and off...

-- 
Martin <http://launchpad.net/~mbp/>