[Qemu-devel] POLL: Support for instrumentation?

[Qemu-devel] POLL: Support for instrumentation?

[Sami Kiminki]

Hi,

We are considering doing some work on instrumentation support for QEMU.
But before rushing into implementation, I wonder if there is more
general interest in this subject, and if so, what kind of
instrumentation would you like to see?

Regards,
Sami Kiminki
Embedded Software Group / Helsinki University of Technology

Re: [Qemu-devel] POLL: Support for instrumentation?

[Luis Useche]

On Thu, Jun 18, 2009 at 05:54:02PM +0300, Sami Kiminki wrote:
> We are considering doing some work on instrumentation support for QEMU.
> But before rushing into implementation, I wonder if there is more
> general interest in this subject, and if so, what kind of
> instrumentation would you like to see?

I have been trying to instrument qemu code to get the traces of memory
accesses. I have seen several threads in this mail list talking about this
topic so probably it will be well received.

Bochs provides an instrumentation utility that you can probably look to
get some ideas.

Best,
-- 
Luis Useche <luis@cs.fiu.edu>
Ph.D. Student
Florida International University

Re: [Qemu-devel] POLL: Support for instrumentation?

[Laurent Desnogues]

On Thu, Jun 18, 2009 at 4:54 PM, Sami Kiminki<sami.kiminki@tkk.fi> wrote:
>
> We are considering doing some work on instrumentation support for QEMU.
> But before rushing into implementation, I wonder if there is more
> general interest in this subject, and if so, what kind of
> instrumentation would you like to see?

I am definitely interested. But again, being the less intrusive
possible is a must from my point of view. Not degrading
performance too much is also a must.

Things I'd like to see:

- instruction traces (including category mix)
- memory traces
- canceled conditional instructions
- taken/non taken branches
- flexible branch prediction
- flexible multi-level cache simulation

Some of these things can be done after execution, provided
enough information is collected at runtime.  By flexible, I mean
an easy to change for instance predictor(s) and cache
geometry.

This is a huge subject with a huge literature :-)

>From an implementation point of view, doing it so that it is
easy to add runtime info collection and post execution
info extraction will be difficult, especially if you want it to
be accepted into mainline.


Laurent