On Thu, Apr 28, 2016 at 10:41:10AM -0700, Eric Dumazet wrote:
> On Thu, 2016-04-28 at 10:23 -0700, Alexei Starovoitov wrote:
> > On Wed, Apr 27, 2016 at 10:25:46PM -0700, Eric Dumazet wrote:
> > > Most of TCP stack assumed it was running from BH handler.
> > >
> > > This is great for most things, as TCP behavior is very sensitive
> > > to scheduling artifacts.
> > >
> > > However, the prequeue and backlog processing are problematic,
> > > as they need to be flushed with BH being blocked.
> > >
> > > To cope with modern needs, TCP sockets have big sk_rcvbuf values,
> > > in the order of 16 MB.
> > > This means that backlog can hold thousands of packets, and things
> > > like TCP coalescing or collapsing on this amount of packets can
> > > lead to insane latency spikes, since BH are blocked for too long.
> > >
> > > It is time to make UDP/TCP stacks preemptible.
> > >
> > > Note that fast path still runs from BH handler.
> >
> > this looks pretty awesome.
>
> Yes, I am pretty excited ;)
>
> > the change will make the backlog run in bh enabled, so that one
> > large flow reciever will not penalize the rest of the system, right?
>
> Not only large flows, but flows with losses/reorders.
> Typically many flows are in this case when a congestion collapse
> happens.
>
>
>
> > but you're saying that prequeue is also expensive, but not touched
> > by this patchset? was it addressed by your eariler patch?
> > Or more work still tbd?
>
> prequeue is handled by "[2/6] tcp: do not block bh during prequeue
> processing"
got it. It applies to both v4 and v6, right?
> Note that I also sent a patch earlier (("tcp: give prequeue mode some
> care")) to control max size of prequeue to 32 packets.
>
> > I'm just trying to understand more about tcp stack.
> >
> Sure ;)
>
> I also have a patch to add scheduling point in sendmsg() (ie : draining
> the backlog if not empty) for each new skb added to the write queue.
>
> Since each skb is about 64KB (with GSO/TSO), it means an application no
> longer will hold the socket lock too long, even when doing a
> write()/sendmsg() of say 8 MB at once ;)
that would be awesome. I think map/reduce type jobs typically
do large sendmsg, so it should help p99 latency.
