Re: Floating point in pango
- From: Soeren Sandmann <sandmann daimi au dk>
- To: mallum openedhand com
- Cc: Behdad Esfahbod <behdad behdad org>, performance-list gnome org
- Subject: Re: Floating point in pango
- Date: 28 Jul 2006 04:08:30 +0200
Matthew Allum <mallum openedhand com> writes:
> But I think the difference here is, and a point you've very well made,
> is sysprof is designed for desktop type hardware - thats not (usually)
> ARM. On ARM we *need* features that oprofile gives us like use of
> hardware foo to lower overhead and the data collection separation.
> Porting sysprof to ARM as to be useful would likely break your original
> design goals.
I wouldn't say that sysprof is designed for desktop type _hardware_,
as much as it is designed for desktop developers. After all, an ARM of
today is as powerful as a desktop not that many years ago, and sysprof
actually goes to quite some length to limit the amount of CPU and
memory it uses.
I don't really see the overhead of using a timer interrupt rather than
a "counter-overflow" interrupt; in fact I would expect that actually
walking the stack would be the thing that dominated, not the handling
of the interrupt. I'll admit though that sysprof at the moment
requires a system call per sample which oprofile doesn't. I should
probably fix that.
Also note that cvs HEAD of sysprof actually does have separation of
collection and analysis. There is a command line tool, sysprof-text,
producing files that can be loaded by the graphical tool.
> Sysprof does make more sense for your average developer, who just wants
> to profile his code with the minimum amount of hassle, but on the
> *desktop*. On embedded, where its a different set of constraints,
> oprofile currently makes more sense.
Feel free to use whatever makes more sense. In the end I care a lot
more about people doing useful optimizations, than I do about what
tools they use.
What I objected to, and still object to, is the idea that oprofile is
"technically better". The features that the oprofile kernel module has
over the sysprof module (such as the ability to use cache miss and
branch misprediction counters) are quite simply not useful for what
sysprof tries to do.
Soren
[
Date Prev][
Date Next] [
Thread Prev][
Thread Next]
[
Thread Index]
[
Date Index]
[
Author Index]