All of lore.kernel.org
 help / color / mirror / Atom feed
* marvell 6190 NAT performance
@ 2019-01-24 20:26 Marek Behun
  2019-01-24 21:31 ` Florian Fainelli
  0 siblings, 1 reply; 5+ messages in thread
From: Marek Behun @ 2019-01-24 20:26 UTC (permalink / raw)
  To: Florian Fainelli; +Cc: Andrew Lunn, netdev, Russell King

Hello,

I am encountering strange performance issue when benchmarking NAT
performance on Armada 3720 with Marvell 88e6190 switch.

Download speed (from internet, via Armada 3720 NAT, via switch to LAN
device) is ~750mbps and the CPU running on 100% (mostly in ksoftirq).
Upload speed is ~250mbps.

When the LAN device is connected to A3720 directly (via SFP), the
speeds are both ~1000mbps.

I realize that packing/unpacking packets with Marvell header for the
switch takes some time, but is such a performance drop expected?

This was tested with 5.0.0-rc2 and also 4.14.

Thank you.

Marek

^ permalink raw reply	[flat|nested] 5+ messages in thread

* Re: marvell 6190 NAT performance
  2019-01-24 20:26 marvell 6190 NAT performance Marek Behun
@ 2019-01-24 21:31 ` Florian Fainelli
  2019-01-25 12:09   ` Marek Behún
  2019-01-29 14:27   ` Marek Behún
  0 siblings, 2 replies; 5+ messages in thread
From: Florian Fainelli @ 2019-01-24 21:31 UTC (permalink / raw)
  To: Marek Behun; +Cc: Andrew Lunn, netdev, Russell King

On 1/24/19 12:26 PM, Marek Behun wrote:
> Hello,
> 
> I am encountering strange performance issue when benchmarking NAT
> performance on Armada 3720 with Marvell 88e6190 switch.
> 
> Download speed (from internet, via Armada 3720 NAT, via switch to LAN
> device) is ~750mbps and the CPU running on 100% (mostly in ksoftirq).
> Upload speed is ~250mbps.
> 
> When the LAN device is connected to A3720 directly (via SFP), the
> speeds are both ~1000mbps.

OK and that presumably uses the second Ethernet MAC on the SoC right?

> 
> I realize that packing/unpacking packets with Marvell header for the
> switch takes some time, but is such a performance drop expected?

If you run perf top/record you would be able to see that pretty quickly,
I would not think that processing of the Marvell DSA tag would incur
such a high penalty though since the packets are already hot in D$ by
the time we get to mangle them for the DSA network devices.

How about pure (non-NAT) IP routing? How about just bridging between WAN
and LAN?

> 
> This was tested with 5.0.0-rc2 and also 4.14.
> 
> Thank you.
> 
> Marek
> 

-- 
Florian

^ permalink raw reply	[flat|nested] 5+ messages in thread

* Re: marvell 6190 NAT performance
  2019-01-24 21:31 ` Florian Fainelli
@ 2019-01-25 12:09   ` Marek Behún
  2019-01-29 14:27   ` Marek Behún
  1 sibling, 0 replies; 5+ messages in thread
From: Marek Behún @ 2019-01-25 12:09 UTC (permalink / raw)
  To: Florian Fainelli; +Cc: Andrew Lunn, netdev, Russell King

On Thu, 24 Jan 2019 13:31:24 -0800
Florian Fainelli <f.fainelli@gmail.com> wrote:

> On 1/24/19 12:26 PM, Marek Behun wrote:
> > Hello,
> > 
> > I am encountering strange performance issue when benchmarking NAT
> > performance on Armada 3720 with Marvell 88e6190 switch.
> > 
> > Download speed (from internet, via Armada 3720 NAT, via switch to
> > LAN device) is ~750mbps and the CPU running on 100% (mostly in
> > ksoftirq). Upload speed is ~250mbps.
> > 
> > When the LAN device is connected to A3720 directly (via SFP), the
> > speeds are both ~1000mbps.  
> 
> OK and that presumably uses the second Ethernet MAC on the SoC right?

SFP port uses the same Ethernet MAC as switch.
eth0 is used for wan, eth1 is either connected to a SFP cage or to a
switch chip.

> > 
> > I realize that packing/unpacking packets with Marvell header for the
> > switch takes some time, but is such a performance drop expected?  
> 
> If you run perf top/record you would be able to see that pretty
> quickly, I would not think that processing of the Marvell DSA tag
> would incur such a high penalty though since the packets are already
> hot in D$ by the time we get to mangle them for the DSA network
> devices.
> 
> How about pure (non-NAT) IP routing? How about just bridging between
> WAN and LAN?

I will try to do various benchmarks and send the results.

Marek


^ permalink raw reply	[flat|nested] 5+ messages in thread

* Re: marvell 6190 NAT performance
  2019-01-24 21:31 ` Florian Fainelli
  2019-01-25 12:09   ` Marek Behún
@ 2019-01-29 14:27   ` Marek Behún
  2019-01-29 14:56     ` Andrew Lunn
  1 sibling, 1 reply; 5+ messages in thread
From: Marek Behún @ 2019-01-29 14:27 UTC (permalink / raw)
  To: Florian Fainelli; +Cc: Andrew Lunn, netdev, Russell King

Hi Florian,
I've made a screenshot of perf top when doing the NAT throughput test
without the switch (which too doesn't work on 1000mbps as I thought,
but on ~680 mbps). What do you think about the result?

http://blackhole.sk/~kabel/tmp/a3700_nat_perf.png

Marek

On Thu, 24 Jan 2019 13:31:24 -0800
Florian Fainelli <f.fainelli@gmail.com> wrote:

> On 1/24/19 12:26 PM, Marek Behun wrote:
> > Hello,
> > 
> > I am encountering strange performance issue when benchmarking NAT
> > performance on Armada 3720 with Marvell 88e6190 switch.
> > 
> > Download speed (from internet, via Armada 3720 NAT, via switch to
> > LAN device) is ~750mbps and the CPU running on 100% (mostly in
> > ksoftirq). Upload speed is ~250mbps.
> > 
> > When the LAN device is connected to A3720 directly (via SFP), the
> > speeds are both ~1000mbps.  
> 
> OK and that presumably uses the second Ethernet MAC on the SoC right?
> 
> > 
> > I realize that packing/unpacking packets with Marvell header for the
> > switch takes some time, but is such a performance drop expected?  
> 
> If you run perf top/record you would be able to see that pretty
> quickly, I would not think that processing of the Marvell DSA tag
> would incur such a high penalty though since the packets are already
> hot in D$ by the time we get to mangle them for the DSA network
> devices.
> 
> How about pure (non-NAT) IP routing? How about just bridging between
> WAN and LAN?
> 
> > 
> > This was tested with 5.0.0-rc2 and also 4.14.
> > 
> > Thank you.
> > 
> > Marek
> >   
> 


^ permalink raw reply	[flat|nested] 5+ messages in thread

* Re: marvell 6190 NAT performance
  2019-01-29 14:27   ` Marek Behún
@ 2019-01-29 14:56     ` Andrew Lunn
  0 siblings, 0 replies; 5+ messages in thread
From: Andrew Lunn @ 2019-01-29 14:56 UTC (permalink / raw)
  To: Marek Behún; +Cc: Florian Fainelli, netdev, Russell King

On Tue, Jan 29, 2019 at 03:27:34PM +0100, Marek Behún wrote:
> Hi Florian,
> I've made a screenshot of perf top when doing the NAT throughput test
> without the switch (which too doesn't work on 1000mbps as I thought,
> but on ~680 mbps). What do you think about the result?
> 
> http://blackhole.sk/~kabel/tmp/a3700_nat_perf.png

Hi Marek

This is plain text, you can just cut/paste it into the email.

What you actually want to do is a side by side comparison of this and
the case where it does go through the switch. What are the big
changes?

	Andrew

^ permalink raw reply	[flat|nested] 5+ messages in thread

end of thread, other threads:[~2019-01-29 14:56 UTC | newest]

Thread overview: 5+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2019-01-24 20:26 marvell 6190 NAT performance Marek Behun
2019-01-24 21:31 ` Florian Fainelli
2019-01-25 12:09   ` Marek Behún
2019-01-29 14:27   ` Marek Behún
2019-01-29 14:56     ` Andrew Lunn

This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.