Changing CC algos on the fly

[tohojo]

Changing CC algos on the fly

[tohojo]

Definitely possible to use test-parameter to set the CC, but it would be
a special-casing.

See also #34 - would this be a way to achieve the comparability?

[dtaht]

hahahaha. great minds think alike. You beat me to thinking about it by 4 hours....

Yes. Although I have generally liked rrul and rrul_be to be non-comparable, also, when I look at a large dataset I tend to look at comparisons within those two.

I guess it also comes down to better being able to select the metadata on which multiple plots can be compared.

[dtaht]

I might argue that CC is special enough to be it's own reserved word or parameter, and default means of handing stuff onto netperf. I would of course like to see httping variants gain support also....

[tohojo]

Dave Täht [email protected] writes:

I guess it also comes down to better being able to select the metadata
on which multiple plots can be compared.

Yes, some kind of definition of what 'compatible' means is definitely
needed. Maybe explicit categorisation in test definitions? "This is a
'download' type flow"?

[tohojo]

Dave Täht [email protected] writes:

I might argue that CC is special enough to be it's own reserved word or
parameter, and default means of handing stuff onto netperf. I would of course
like to see httping variants gain support also....

Well I don't mind special-casing it if necessary, but if we can achieve
more with less code I'm happier ;)

[dtaht]

well, the simplest possible thing, adding this to netperf_definitions.inc

CC=TEST_PARAMETERS.get('CC','cubic')

and this to tcp_1up

DESCRIPTION="Single " + CC + " TCP upload stream w/ping"

didn't take the override on the command line, yielding cubic everytime.

I think your intent was that in the gui that TEST_PARAMETERS pull from the metadata on output...

and from the command line on the test itself.

[dtaht]

And I always worry when I see no comma here for some reason between cubic,cubic. Have always kind of figured that was messing up the other CS1,CS1 stuff. But I suppose it is just an artifact of ps....

/usr/local/bin/netperf -P 0 -v 0 -D -0.20 -4 -H netperf-west.bufferbloat.net -t TCP_STREAM -l 60 -f m -- -K cubic cubic

[dtaht]

Assuming I got the first part right and wasn't so dumb about how to extract the TEST_PARAMETERS, this does stick it on the command line. S

find_netperf("TCP_STREAM", LENGTH, HOST, extra_args="-- -K " + CC + "," + CC)

[tohojo]

Dave Täht [email protected] writes:

Assuming I got the first part right and wasn't so dumb about how to extract the
TEST_PARAMETERS, this does stick it on the command line. S

find_netperf("TCP_STREAM", LENGTH, HOST, extra_args="-- -K " + CC + "," + CC)

Well, it would probably be better to teach find_netperf a CC parameter...

[dtaht]

I am still stuck at:

CC=TEST_PARAMETERS.get('CC','cubic')

My assumption was that the first arg would pull from the environment, the second be the default if not there.

[tohojo]

That assumption is correct. However, the keys are case folded to lowercase when taken from the command line, so you'll want

CC=TEST_PARAMETERS.get('cc','cubic')

[dtaht]

OK, that drives the test right. And does the right thing on single plots of each.

but it was my hope that TEST_PARAMETERS was locally scoped, so by combining reno vs cubic plots like this (modifying tcp_1up right now), I would get labels that made sense. Instead it pulls the "CC" variable from the first file pulled from for combinatory plots.

(and arguably error checking for the actual availability of the CC algo in the local system is good, which I'll go do next. I picked the wrong week to try to learn more python)

Maybe a call to "eval"? self->something?

PLOTS = o([
('totals',
{'description': 'Bandwidth and ping plot',
'type': 'timeseries',
'dual_axes': True,
'series': [{'data': 'TCP upload',
'label': TEST_PARAMETERS.get('cc','cubic')

• ' Upload'},

...

[dtaht]

My overall thought now is to add a new test: "tcp", with parameters to vary the cc,number of flows,and hosts, and controlling the direction on the command line as well, combined with an isochronous measurement flow and/or the existing udp+ping measurement flows.

cc=cubic,westwood,etc,etc
flows=n
hosts=x,y,z,q,etc

[tohojo]

Well, right now the test name is used for gauging whether two tests can
be plotted together (but this is a limitation in itself, see #34).

Some thought would have to be put into how to manage multiple tests that
are "the same" but in reality can vary wildly.

[tohojo]

The congestion control algorithm can be set with the tcp_cong_control test parameter.