Libguestfs - Mar 2023 - [PATCH 1/1] nbd/server: push pending frames after sending reply

On Fri, Mar 24, 2023 at 11:47:20AM +0100, Florian Westphal
wrote:
> qemu-nbd doesn't set TCP_NODELAY on the tcp socket.
> 
> Kernel waits for more data and avoids transmission of small packets.
> Without TLS this is barely noticeable, but with TLS this really shows.
> 
> Booting a VM via qemu-nbd on localhost (with tls) takes more than
> 2 minutes on my system.  tcpdump shows frequent wait periods, where no
> packets get sent for a 40ms period.
Thank you for this analysis.
> 
> Add explicit (un)corking when processing (and responding to) requests.
> "TCP_CORK, &zero" after earlier "CORK, &one"
will flush pending data.
> 
> VM Boot time:
> main:    no tls:  23s, with tls: 2m45s
> patched: no tls:  14s, with tls: 15s
> 
> VM Boot time, qemu-nbd via network (same lan):
> main:    no tls:  18s, with tls: 1m50s
> patched: no tls:  17s, with tls: 18s
And the timings bear proof that it matters.
> 
> Future optimization: if we could detect if there is another pending
> request we could defer the uncork operation because more data would be
> appended.
nbdkit and libnbd do this with the MSG_MORE flag (plaintext) and TLS
corking (tls); when building up a message to the other side, a flag is
set any time we know we are likely to send more data very shortly.

nbdkit wraps it under a flag SEND_MORE, which applies to both plaintext:
https://gitlab.com/nbdkit/nbdkit/-/blob/master/server/connections.c#L415
and to TLS connections:
https://gitlab.com/nbdkit/nbdkit/-/blob/master/server/crypto.c#L396

while libnbd uses MSG_MORE a bit more directly for the same purpose
for plaintext, but isn't (yet) doing TLS corking:
https://gitlab.com/nbdkit/libnbd/-/blob/master/generator/states-issue-command.c#L53
https://gitlab.com/nbdkit/libnbd/-/blob/master/lib/internal.h#L57

I would love to see a future patch to qio_channel code to support
MSG_MORE in the same way as nbdkit is using its SEND_MORE flag, as the
caller often has more info on whether it is sending a short prefix or
is done with a conceptual message and ready to uncork, and where the
use of a flag can be more efficient than separate passes through
cork/uncork calls.  But even your initial work at properly corking is
a good step in the right direction.

And surprisingly, qemu IS using corking on the client side:
https://gitlab.com/qemu-project/qemu/-/blob/master/block/nbd.c#L525
just not on the server side, before your patch.
> 
> Signed-off-by: Florian Westphal <fw at strlen.de>
> ---
>  nbd/server.c | 3 +++
>  1 file changed, 3 insertions(+)
> 
> diff --git a/nbd/server.c b/nbd/server.c
> index a4750e41880a..848836d41405 100644
> --- a/nbd/server.c
> +++ b/nbd/server.c
> @@ -2667,6 +2667,8 @@ static coroutine_fn void nbd_trip(void *opaque)
>          goto disconnect;
>      }
>  
> +    qio_channel_set_cork(client->ioc, true);
> +
>      if (ret < 0) {
>          /* It wasn't -EIO, so, according to nbd_co_receive_request()
>           * semantics, we should return the error to the client. */
> @@ -2692,6 +2694,7 @@ static coroutine_fn void nbd_trip(void *opaque)
>          goto disconnect;
>      }
>  
> +    qio_channel_set_cork(client->ioc, false);
Reviewed-by: Eric Blake <eblake at redhat.com>
>  done:
>      nbd_request_put(req);
>      nbd_client_put(client);
> -- 
> 2.39.2
> 
-- 
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3266
Virtualization:  qemu.org | libvirt.org

On Fri, Mar 24, 2023 at 02:41:20PM -0500, Eric Blake
wrote:
> On Fri, Mar 24, 2023 at 11:47:20AM +0100, Florian Westphal wrote:
> > qemu-nbd doesn't set TCP_NODELAY on the tcp socket.
Replying to myself, WHY aren't we setting TCP_NODELAY on the socket?
> 
> And surprisingly, qemu IS using corking on the client side:
> https://gitlab.com/qemu-project/qemu/-/blob/master/block/nbd.c#L525
> just not on the server side, before your patch.
Corking matters more when TCP_NODELAY is enabled.  The entire reason
Nagle's algorithm exists (and is default on unless you enable
TCP_NODELAY) is that the benefits of merging smaller piecemeal packets
into larger traffic is a lot easier to do in a common location for
code that isn't super-sensitive to latency and message boundaries.
But once you get to the point where corking or MSG_MORE makes a
difference, then it is obvious you know your message boundaries, and
will benefit from TCP_NODELAY, at the expense of potentially more
network traffic overhead.  One more code search, and I find that we
use TCP_NODELAY in all of:

qemu client:
https://gitlab.com/qemu-project/qemu/-/blob/master/nbd/client-connection.c#L143
nbdkit: https://gitlab.com/nbdkit/nbdkit/-/blob/master/server/sockets.c#L430
libnbd:
https://gitlab.com/nbdkit/libnbd/-/blob/master/generator/states-connect.c#L41

so I think we _should_ be calling qio_channel_set_delay(false) for
qemu-nbd as well.  That doesn't negate your patch, but rather argues
that we can go for even better performance with TCP_NODELAY also
turned on.

-- 
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3266
Virtualization:  qemu.org | libvirt.org