Re: [Qemu-devel] [PATCH v3 1/2] Add bdrv_aio_multiwrite

[Christoph Hellwig <hch@lst.de>]

On Fri, Sep 11, 2009 at 08:29:09AM +0200, Kevin Wolf wrote:
> Am 11.09.2009 00:44, schrieb Christoph Hellwig:
> > One thing that concerns me here is that we keep adding more memory
> > allocations to the I/O path.  At least on fast SSDs even kernel memory
> > allocations are a performance problem and they're much faster than
> > userspace ones.  
> 
> In the non-merging case we have one additional allocation, the one for
> mcb. Maybe we could do something like the AIOPool here to avoid some
> mallocs.
> 
> In the merging case we need to allocate new IO vectors. I don't see a
> way around this, though, and I hope that the benefit of a saved write
> request outweighs the malloc costs. If not, it probably was a bad idea
> to accept your suggestion to make it generic code instead of just
> merging requests for qcow2.

Right now I'm just thinking about the potential issues.  The problem
is that the trade offs maybe be very different for different scenarios.
For a normal disk which is by far the most common option (and will be
for a long time) the merging is a clear benefit.  For SSDs we might have
to do different optimizations.  If it's not a big issue avoiding too
many memory allocations and rather doing fewer larger ones is in general
a good thing, though.