Re: [Qemu-devel] [PATCH] Add a memory barrier to DMA functions

[Anthony Liguori <anthony@codemonkey.ws>]

On 05/22/2012 07:03 AM, Michael S. Tsirkin wrote:
> On Tue, May 22, 2012 at 09:41:41PM +1000, Benjamin Herrenschmidt wrote:
>> On Tue, 2012-05-22 at 14:14 +0300, Michael S. Tsirkin wrote:
>>>> The baseline is the laptop without kvm talking to the server. The
>>>> TCP_STREAM test results are:
>>>
>>> It's not a good test. The thing most affecting throughput results is
>>> how
>>> much CPU does you guest get. So as a minumum you need to measure CPU
>>> utilization on the host and divide by that.
>>
>> The simple fact that we don't reach the baseline while in qemu seems to
>> be a reasonably good indication that we tend to be CPU bound already so
>> it's not -that- relevant. It would be if we were saturating the network.
>>
>> But yes, I can try to do more tests tomorrow, it would be nice if you
>> could contribute a proper test protocol (or even test on some machines)
>> since you seem to be familiar with those measurements (and I have a very
>> limited access to x86 gear ... basically just my laptop).
>>
>> Cheers,
>> Ben.
>
> I have a deja vu. Amos sent perf results when you argued about
> exactly the same issue in guest virtio. Delta was small but
> measureable. At the moment I have no free time or free hardware
> to redo the same work all over again. It's a well known fact that
> actual memory barrier is slow on x86 CPUs. You can't see
> it with network on your laptop? Write a microbenchmark.

The latest patch doesn't put a barrier in map().  I have an extremely hard time 
believing anything that uses _rw is going to be performance sensitive.

There is a correctness problem here.  I think it's important that we fix that 
and then we can focus on improving performance later.

Regards,

Anthony Liguori

>