Herbert Xu wrote:
On Thu, Feb 05, 2009 at 02:37:07PM +0200, Avi Kivity wrote:
I believe that copyless networking is absolutely essential.
I used to think it was important, but I'm now of the opinion
that it's quite useless for virtualisation as it stands.
For transmit, copyless is needed to properly support sendfile() type
workloads - http/ftp/nfs serving. These are usually high-bandwidth,
cache-cold workloads where a copy is most expensive.
This is totally true for baremetal, but useless for virtualisation
right now because the block layer is not zero-copy. That is, the
data is going to be cache hot anyway so zero-copy networking doesn't
buy you much at all.
The guest's block layer is copyless. The host block layer is -><- this
far from being copyless -- all we need is preadv()/pwritev() or to
replace our thread pool implementation in qemu with linux-aio.
Everything else is copyless.
Since we are actively working on this, expect this limitation to
disappear soon.
(even if it doesn't, the effect of block layer copies is multiplied by
the cache miss percentage which can be quite low for many workloads; but
again, we're not bulding on that)
Please also recall that for the time being, block speeds are
way slower than network speeds. So the really interesting case
is actually network-to-network transfers. Again due to the
RX copy this is going to be cache hot.
Block speeds are not way slower. We're at 4Gb/sec for Fibre and 10Gb/s
for networking. With dual channels or a decent cache hit rate they're
evenly matched.
For receive, the guest will almost always do an additional copy, but it
will most likely do the copy from another cpu. Xen netchannel2
That's what we should strive to avoid. The best scenario with
modern 10GbE NICs is to stay on one CPU if at all possible. The
NIC will pick a CPU when it delivers the packet into one of the
RX queues and we should stick with it for as long as possible.
So what I'd like to see next in virtualised networking is virtual
multiqueue support in guest drivers. No I'm not talking about
making one or more of the physical RX/TX queues available to the
guest (aka passthrough), but actually turning something like the
virtio-net interface into a multiqueue interface.
I support this, but it should be in addition to copylessness, not on its
own.
- many guests will not support multiqueue
- for some threaded workloads, you cannot predict where the final read()
will come from; this renders multiqueue ineffective for keeping cache
locality
- usually you want virtio to transfer large amounts of data; but if you
want your copies to be cache-hot, you need to limit transfers to half
the cache size (a quarter if hyperthreading); this limits virtio
effectiveness
--
I have a truly marvellous patch that fixes the bug which this
signature is too narrow to contain.
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
