[Qemu-devel] A question regarding QEMU.

[Qemu-devel] A question regarding QEMU.

[Chung Hwan Kim]

Hi! My name is Chunghwan. I am a Masters student working the Flux
Systems Research Group, the University of Utah, USA.

I have a question regarding QEMU. First, let me introduce what I am
doing related with QEMU.

I and two other students have formed up a team for a project called
"Accelerating Dynamic Binary Translation with the GPUs". As the name of
the project suggests our main idea is to parallelize Dynamic Binary
Translation (DBT) process and speed it up with GPUs using the NVIDIA
CUDA library. 

As far as we know the target of DBT can be - a process (OS process) or
a virtual machine. When the target is a VM we do not know where
to begin translation because the execution flow of a VM is
nondeterministic.

We would like to know whether you think it is possible for us to
parallelize the DBT in QEMU. Moreover, we hope you can give
us advice on figuring out the control flow of VM execution.


-- 
Sincerely,

Chung Hwan Kim
School of Computing and Flux Research Group
University of Utah
chunghwn@cs.utah.edu
www.cs.utah.edu/~chunghwn

Re: [Qemu-devel] A question regarding QEMU.

[Lluís]

Chung Hwan Kim writes:

> I and two other students have formed up a team for a project called
> "Accelerating Dynamic Binary Translation with the GPUs". As the name of
> the project suggests our main idea is to parallelize Dynamic Binary
> Translation (DBT) process and speed it up with GPUs using the NVIDIA
> CUDA library.

AFAIK, DBT is a fairly control flow intensive code, so you'll probably
run into lots of branch divergence problems, so that performance will
suffer a lot, even if you use instruction template tables (like in
qemu's PPC target).

Nonetheless, I think new fermi models have less problems with that, but
it's still an architecture thought for control-flow-homogeneous parallel
code.

In any case, I'm not sure what is the real cost of translation related
to execution, it all depends on the kind of applications you're running;
but the computation on the GPU better have a huge speedup compared to
the current approach, or otherwise the data transfers to/from the GPU
will dominate the cost, specially if they're small transfers.


Lluis

--
 "And it's much the same thing with knowledge, for whenever you learn
 something new, the whole world becomes that much richer."
 -- The Princess of Pure Reason, as told by Norton Juster in The Phantom
 Tollbooth