Date:   Thu, 2 Aug 2018 01:27:13 +0300
From:   "Michael S. Tsirkin" <mst@...hat.com>
To:     Jason Wang <jasowang@...hat.com>
Cc:     Jon Olson <jonolson@...gle.com>, willemdebruijn.kernel@...il.com,
        caleb.raitto@...il.com, davem@...emloft.net,
        netdev@...r.kernel.org, Caleb Raitto <caraitto@...gle.com>
Subject: Re: [PATCH net-next] virtio_net: force_napi_tx module param.

On Mon, Jul 30, 2018 at 02:06:50PM +0800, Jason Wang wrote:
> 
> 
> On 2018年07月25日 08:17, Jon Olson wrote:
> > On Tue, Jul 24, 2018 at 3:46 PM Michael S. Tsirkin <mst@...hat.com> wrote:
> > > On Tue, Jul 24, 2018 at 06:31:54PM -0400, Willem de Bruijn wrote:
> > > > On Tue, Jul 24, 2018 at 6:23 PM Michael S. Tsirkin <mst@...hat.com> wrote:
> > > > > On Tue, Jul 24, 2018 at 04:52:53PM -0400, Willem de Bruijn wrote:
> > > > > > >From the above linked patch, I understand that there are yet
> > > > > > other special cases in production, such as a hard cap on #tx queues to
> > > > > > 32 regardless of number of vcpus.
> > > > > I don't think upstream kernels have this limit - we can
> > > > > now use vmalloc for higher number of queues.
> > > > Yes. that patch* mentioned it as a google compute engine imposed
> > > > limit. It is exactly such cloud provider imposed rules that I'm
> > > > concerned about working around in upstream drivers.
> > > > 
> > > > * for reference, I mean https://patchwork.ozlabs.org/patch/725249/
> > > Yea. Why does GCE do it btw?
> > There are a few reasons for the limit, some historical, some current.
> > 
> > Historically we did this because of a kernel limit on the number of
> > TAP queues (in Montreal I thought this limit was 32). To my chagrin,
> > the limit upstream at the time we did it was actually eight. We had
> > increased the limit from eight to 32 internally, and it appears in
> > upstream it has subsequently increased upstream to 256. We no longer
> > use TAP for networking, so that constraint no longer applies for us,
> > but when looking at removing/raising the limit we discovered no
> > workloads that clearly benefited from lifting it, and it also placed
> > more pressure on our virtual networking stack particularly on the Tx
> > side. We left it as-is.
> > 
> > In terms of current reasons there are really two. One is memory usage.
> > As you know, virtio-net uses rx/tx pairs, so there's an expectation
> > that the guest will have an Rx queue for every Tx queue. We run our
> > individual virtqueues fairly deep (4096 entries) to give guests a wide
> > time window for re-posting Rx buffers and avoiding starvation on
> > packet delivery. Filling an Rx vring with max-sized mergeable buffers
> > (4096 bytes) is 16MB of GFP_ATOMIC allocations. At 32 queues this can
> > be up to 512MB of memory posted for network buffers. Scaling this to
> > the largest VM GCE offers today (160 VCPUs -- n1-ultramem-160) keeping
> > all of the Rx rings full would (in the large average Rx packet size
> > case) consume up to 2.5 GB(!) of guest RAM. Now, those VMs have 3.8T
> > of RAM available, but I don't believe we've observed a situation where
> > they would have benefited from having 2.5 gigs of buffers posted for
> > incoming network traffic :)
> 
> We can work to have async txq and rxq instead of paris if there's a strong
> requirement.

I think the reason we don't is because RX queueing is programmed by TX
packets.  It might make sense if we support RX queueing policy that
isn't dependent on TX.


> > 
> > The second reason is interrupt related -- as I mentioned above, we
> > have found no workloads that clearly benefit from so many queues, but
> > we have found workloads that degrade. In particular workloads that do
> > a lot of small packet processing but which aren't extremely latency
> > sensitive can achieve higher PPS by taking fewer interrupt across
> > fewer VCPUs due to better batching (this also incurs higher latency,
> > but at the limit the "busy" cores end up suppressing most interrupts
> > and spending most of their cycles farming out work). Memcache is a
> > good example here, particularly if the latency targets for request
> > completion are in the ~milliseconds range (rather than the
> > microseconds we typically strive for with TCP_RR-style workloads).
> > 
> > All of that said, we haven't been forthcoming with data (and
> > unfortunately I don't have it handy in a useful form, otherwise I'd
> > simply post it here), so I understand the hesitation to simply run
> > with napi_tx across the board. As Willem said, this patch seemed like
> > the least disruptive way to allow us to continue down the road of
> > "universal" NAPI Tx and to hopefully get data across enough workloads
> > (with VMs small, large, and absurdly large :) to present a compelling
> > argument in one direction or another. As far as I know there aren't
> > currently any NAPI related ethtool commands (based on a quick perusal
> > of ethtool.h)
> 
> As I suggest before, maybe we can (ab)use tx-frames-irq.
> 
> Thanks
> 
> > -- it seems like it would be fairly involved/heavyweight
> > to plumb one solely for this unless NAPI Tx is something many users
> > will want to tune (and for which other drivers would support tuning).
> > 
> > --
> > Jon Olson

Powered by blists - more mailing lists