Re: [Xen-users] xen network para-virtualization

[张伟 <zhangwqh@xxxxxxx>]

Thank you very much, Ian! See below. 

在 2014-03-09 22:15:35，"Ian Campbell" <ian.campbell@xxxxxxxxxx> 写道：
>On Sun, 2014-03-09 at 20:48 +0800, 张伟 wrote:
>> Hi all, 
>> 
>> 
>> I have a question about xen network para-virtualization. From the web
>> information, it seems that xen supports copy and page flipping for
>> packet transmission between netback and netfront. Does anybody know
>> which method is efficient?  Why? At first, I think page flipping is
>> more efficient than copy. However, when I read the linux 3.7 code, I
>> found that it only supports copy. If page flipping is faster than
>> copy, why the newer kernel version  does not support it. So I think my
>> understanding has some problem. Does anybody give me some
>> instructions? Thank you in advance!
>
>There are actually three modes here, copy (both directions), flipping
>(guest rx) and mapping (guest tx).
>
>IIRC very old kernels (pre 2.6.18) supported page flipping for traffic
>going back->front but IIRC the TLB shootdown overheads we re so high
>that copy was found to be more efficient. (NB: this was before my time
>so I am not 100% sure of the specifics).

Can you give some explanation why page flipping can bring TLB shootdown overheads high?
>
>The 2.6.18 ports also supported mapping for traffic going from
>back->front. However this required infrastructure which was very
>difficult to upstream and so a version which used copying was upstreamed
>into the mainline kernel instead. Zoltan Kiss has been working recently
>on reintroducing mapping mode to the upstream netback (see xen-devel@
>archives).
What do you mean 2.6.18 port? Which version? Do you think why does not use mapping for rx, only tx supports mapping? Why it is difficult to upstream? Which parts obstacle the implementation in upstreamed version?

>At one point there was some speculation that copying might be faster on
>newer processors which are more optimised for such cases. 

Before the tests, why did you sepculation that copying might be faster, which processors new technology help to make it faster?

That doesn't
>seem to be the case in practice though, at least not according to the
>numbers Zoltan has been reporting.
>
>Ian.
>
>
>
>_______________________________________________
>Xen-users mailing list
>Xen-users@xxxxxxxxxxxxx
>http://lists.xen.org/xen-users

_______________________________________________
Xen-users mailing list
Xen-users@xxxxxxxxxxxxx
http://lists.xen.org/xen-users