Last round of Scale testing

[roger peppe]

[reposted without the big attachment]

On 7 November 2013 13:18, John Arbash Meinel <john at arbash-meinel.com> wrote:
> There is still the issue of what the process was doing when it hit 0%
> CPU. I uploaded the traceback (xz compressed) to Ubuntu One, and it
> should be here:
> * sigquit.log.xz (2.7 MB) hosted on Ubuntu One:
> http://ubuntuone.com/6vKrM07H7BTHUa9Xqvy16V

I was interested to try to see what was going on here, so
I hacked up a quick analysis program that puts all
the stack traces into a call graph summarising the overall
situation.

It makes for an interesting snapshot of a juju system in action,
although I haven't gained any particular insights
from it yet.

I've attached the output svg (suitable for web browser viewing) and
the program I used to generate it.

The nodes of the graph are functions; the count for each node
is the number of goroutines in the stack traces that call that function;
the count for each edge is the number of goroutines that have
made the call between the two connected functions.
The function that created a goroutine is included too.

All cudos to go tool pprof for the svg stuff.

https://s3.amazonaws.com/rogpeppe-scratch/callstack.svg
-------------- next part --------------
A non-text attachment was scrubbed...
Name: stackgraph.go
Type: text/x-go
Size: 12737 bytes
Desc: not available
URL: <https://lists.ubuntu.com/archives/juju-dev/attachments/20131107/f75e9ef6/attachment-0001.bin>