Jeff Roberson, the primary developer of FreeBSD's revived ULE [1] scheduler [story [2]], has committed a new a python/tkinter tool, schedgraph [3], to the freebsd -current tree. Schedgraph will assist with scheduler testing and refinement as well as help developers study application load and corresponding system behavior. In the initial commit message, Jeff gives a simple description:
"Schedgraph takes input from files produces by ktrdump -ct when KTR_SCHED is compiled into the kernel. The output represents the states of each thread with colored line segments as well as colored points for non-state scheduler events. Each line segment and point is clickable to obtain extra detail."
Jeff includes a screenshot and sample data. Robert Watson follows up with a link to pointers on getting KTR [4] working. Jeff has also been very busy committing more fixes [5] to the ULE scheduler, a couple of which solve long standing bugs. Read on for details.
From: Jeff Roberson [email blocked]
To: freebsd-current
Subject: schedgraph.py
Date: 2004-12-26 0:31:19
I took a break from working on VFS to implement a tool that will help me
further refine the ULE scheduler. This may also be interesting in
understanding application behavior under load, analyzing lock contention
and preemption in the kernel, etc.
To use the tool, you will need to define KTR_SCHED in KTR_COMPILE and
KTR_MASK. I'd also bump entires up to 32768 or larger so you can grab a
few seconds of data. Run your workload, and then capture the data with
'ktrdump -ct > ktr.out'. Then you simply run python schedgraph.py
ktr.out. This requires a recent version of python and
ports/x11-toolkits/py-tkinter.
Here's a screenshot from a recent run:
http://www.chesapeake.net/~jroberson/schedgraph.jpg [6]
I also have some sample data at
http://www.chesapeake.net/~jroberson/smp.out.gz [7] if you want to play with
the tool without capturing data.
The configuration page acts as a legend so you can understand the colors.
The least obvious feature of the display is that the background color
changes according to the cpu that the thread is executing on. In the
screenshot I posted, light grey is cpu 0 and dark grey is cpu 1.
You can also click on any event for greater detail. For events that exist
on two threads, you may click on the corrisponding thread's name in the
event popup to change to that event.
Feedback welcome, patches for new features are even more welcome.
Cheers,
Jeff
---------- Forwarded message ----------
Date: Sun, 26 Dec 2004 00:13:07 +0000 (UTC)
From: Jeff Roberson [email blocked]
To: src-committers, cvs-src, cvs-all
Subject: cvs commit: src/tools/sched schedgraph.py
jeff 2004-12-26 00:13:07 UTC
FreeBSD src repository
Added files:
tools/sched schedgraph.py
Log:
- Add 'schedgraph' a scheduler trace visualization tool written with
python and tkinter. Schedgraph takes input from files produces by
ktrdump -ct when KTR_SCHED is compiled into the kernel. The output
represents the states of each thread with colored line segments as well
as colored points for non-state scheduler events. Each line segment and
point is clickable to obtain extra detail.
Revision Changes Path
1.1 +1209 -0 src/tools/sched/schedgraph.py (new)
From: Julian Elischer [email blocked]
Subject: Re: schedgraph.py
Date: 2004-12-26 6:30:02
Jeff Roberson wrote:
> I took a break from working on VFS to implement a tool that will help me
> further refine the ULE scheduler. This may also be interesting in
> understanding application behavior under load, analyzing lock contention
> and preemption in the kernel, etc.
>
[...]
robert watson has been doing some work in a similar (but not as detailed)
vein. he has some code that produces some quite nice graphs
of threads and cpus using KTR.
From: Robert Watson [email blocked]
Subject: Re: schedgraph.py
Date: 2004-12-26 11:02:48
On Sat, 25 Dec 2004, Jeff Roberson wrote:
> To use the tool, you will need to define KTR_SCHED in KTR_COMPILE and
> KTR_MASK. I'd also bump entires up to 32768 or larger so you can grab a
> few seconds of data. Run your workload, and then capture the data with
> 'ktrdump -ct > ktr.out'. Then you simply run python schedgraph.py
> ktr.out. This requires a recent version of python and
> ports/x11-toolkits/py-tkinter.
Great!
For those who need a little more hand-holding getting KTR running, here's
a URL to try:
http://www.watson.org/~robert/freebsd/netperf/ktr/ [8]
It's been my hope people would start producing more post-processing tools
-- KTR can collect some really great data that's just sitting there
waiting to be mined. I'd be interested in seeing post-processing tools
for locking as well. This looks like a great tool that will be really
helpful in understanding behavior and performance.
Thanks!
Robert N M Watson
From: Jeff Roberson [email blocked]
Subject: Re: schedgraph.py
Date: 2004-12-26 18:17:41
On Sun, 26 Dec 2004, Robert Watson wrote:
>
> On Sat, 25 Dec 2004, Jeff Roberson wrote:
>
> > To use the tool, you will need to define KTR_SCHED in KTR_COMPILE and
> > KTR_MASK. I'd also bump entires up to 32768 or larger so you can grab a
> > few seconds of data. Run your workload, and then capture the data with
> > 'ktrdump -ct > ktr.out'. Then you simply run python schedgraph.py
> > ktr.out. This requires a recent version of python and
> > ports/x11-toolkits/py-tkinter.
>
> Great!
>
> For those who need a little more hand-holding getting KTR running, here's
> a URL to try:
>
> http://www.watson.org/~robert/freebsd/netperf/ktr/ [9]
>
> It's been my hope people would start producing more post-processing tools
> -- KTR can collect some really great data that's just sitting there
> waiting to be mined. I'd be interested in seeing post-processing tools
> for locking as well. This looks like a great tool that will be really
> helpful in understanding behavior and performance.
Well, if you want to display contention, you can filter on that using the
configuration menu, and display only contested locks. This tool could be
easily extended to add events for picking up and droping mutexes as well.
It would only require a new KTR and a regexp to match it.
- Archive of the Above Thread [10]