Re: Routing 10-40 Mpps on OpenBSD

2016-09-13 Thread Martin Pieuchot
On 11/09/16(Sun) 18:04, K K wrote:
> [...]
> > There is a lot of ongoing work in this area, OpenBSD doesn't claim to
> > be the performance leader today.
> 
> What is the take of OpenBSD developers on this?
> Are they any plans?
> 
> Many options seems available, but I have no idea how they could be
> integrated in OpenBSD. I now clearly nothing of proper software
> development.
> 
> - DPDK (now BSD licensed)
> - NETMAP/FW

The plan is to keep being a general purpose operating system with a
kernel abstracting the hardware. 

With that in mind, we're doing our best to modernize the network stack
in order to improve the performance on today's hardware.

There's a lot to do, and what's missing is people working with us.  For
example I send diff daily to improve the network stack and make it
faster.  If you can test them, you're helping *a lot*.  Sadly, very few
people test, fewer comment, and even fewer write diffs.

So as you said, many options are available.  But who's going to do the
work?

PS: Here's the last diff I sent that needs testing:
https://marc.info/?l=openbsd-tech&m=147368131715529&w=2



Re: Routing 10-40 Mpps on OpenBSD

2016-09-12 Thread Stefan Sperling
On Sun, Sep 11, 2016 at 05:46:48PM +, K K wrote:
> Chelsio NIcs: Chelsio T540-CR (although not sure there is an OpenBSD driver)

There is no driver for these cards. There used to be a work-in-progress
driver but it was never finished and hence deleted one year ago.

CVSROOT:/cvs
Module name:src
Changes by: clau...@cvs.openbsd.org 2015/09/11 07:35:15

Modified files:
sys/dev/pci: files.pci 
Removed files:
sys/dev/pci: if_che.c 

Log message:
Remove the unfinished che(4) driver for Chelsio 10G cards. I lost interest
fixing this long ago.



Re: Routing 10-40 Mpps on OpenBSD

2016-09-11 Thread Chris Cappuccio
K K [kk...@outlook.com] wrote:

> I thought Intel, but I speak out of impressions, not backed by any facts.
> 

David Gwynne who is working on the Myricom driver recommends the intel card
if that helps

> What is the take of OpenBSD developers on this?
> Are they any plans?
> 

There's a lot of work going into multi-threading the stack right now.
If you read www.openbsd.org/papers/ and undeadly.org you can keep up
with some of the documented progress. 

I don't think anyone is using OpenBSD at 10Mpps on a box today, or if 
they are, that's the upper limit in a configuration without pf. I'd say
that 10Mpps - 40Mpps is a bit past the "typical small ISP".

> Many options seems available, but I have no idea how they could be
> integrated in OpenBSD. I now clearly nothing of proper software
> development.
> 
> - DPDK (now BSD licensed)
> - NETMAP/FW
> 

The general consensus goes against these types of tools at the
moment. I think people want to get the network stack right, first,
before making it hot-pluggable...

Chris



Re: Routing 10-40 Mpps on OpenBSD

2016-09-11 Thread K K
> I think Intel and Myricom are going to be the best-supported 10GbE on

> OpenBSD at the moment.

I thought Intel, but I speak out of impressions, not backed by any facts.

> The best performance today will be with a processor that packs a lot
> of punch into a smaller number of cores. I'm using Xeon E5-1630 v3
> right now. The E5-2xxx series tend to have more cores at lower clock
> speeds. They make more sense on a regular server.

Also came to this conclusion when I picked E5-2697v2.

> There is a lot of ongoing work in this area, OpenBSD doesn't claim to
> be the performance leader today.

What is the take of OpenBSD developers on this?
Are they any plans?

Many options seems available, but I have no idea how they could be
integrated in OpenBSD. I now clearly nothing of proper software
development.

- DPDK (now BSD licensed)
- NETMAP/FW

> Chris

Thank you for your insights.



Re: Routing 10-40 Mpps on OpenBSD

2016-09-11 Thread Hrvoje Popovski
On 11.9.2016. 19:17, K wrote:
> All,
> 
> This message is a call for people who are interested to benchmark commodity
> hardware with the goal of pushing as much PPS as possible through OpenBSD.
> The initial target is to reach 10 Mpps at 64 bytes (or more precisely 84
> bytes with interpacket gap) and if the experiment proves to be successful,
> we would then aim at 40+ Mpps.
> 
> The ultimate goal of this experiment is to build and share with the
> community a recognized hardware configuration that provides a good ground
> for real-world traffic at a typical small ISP.
> 
> We couldn't find such information online. In our case, the final setup
> would be two routers, each with two 10 Gbps uplink to upstreams Internet
> providers and an OSPF and iBGP connection between them. The software
> stack would be based on OpenBSD, OpenBGPD and OpenOSPFD. There is no
> commercial idea around the finding of this experiment.
> 
> While our budget is not unlimited and privately funded (by individuals),
> we are open to hear what hardware specifications people on this list
> would be interested to see. At the moment, we aim for this:
> 
> CPUs: Intel Xeon CPU E5-2697v2, E5-2667v2, E5-2680v3, E5-2640v3
> Intel NICs: Intel 82599ES, X520, X540-{T1/T2/AT2}, 85595, 82598,
> AF/82598, AT/82598, EB/82599, EB/82599 EN
> Chelsio NIcs: Chelsio T540-CR (although not sure there is an OpenBSD driver)
> 
> If you consider other hardware options, please feel free to reply and let us 
> know.
> We surely will not be testing all these configurations, we will most likely 
> pick on
> CPU from the list and 2-3 NICs from the list as well. This experiment might 
> be also
> taken to FreeBSD for comparison. If necessary, we consider sending this
> configuration in a test center with Spirent hardware to validate this.
> 
> Feedbacks, questions, remarks, doubts, irony, are all welcome :-)
> 
> Cheers.
> 

Hi,

if you are optimist like me buy 2 socket box with intel 82599 cards
and with more than 200MB of RAM which is enough for one full BGP feed :)

At first i would buy one 8-core CPU with higher GHz as i can, and when,
and this is optimistic part :), openbsd gets multiqueue ix stuff and RSS
on top of it i would buy second 8-core CPU because it seems that 82599
is having 16 RSS queues.

For now you can get max 1Mpps with only plain routing without any pseudo
interfaces or pf.



Re: Routing 10-40 Mpps on OpenBSD

2016-09-11 Thread Peter N. M. Hansteen
On 09/11/16 19:46, K K wrote:
> // Previous email bounced, so I resend it. Sorry for duplicate //

Just curious, if you look at the bounce, would that be a
DMARC-worshipper failing to understand mailing list mail?

I'm researching what will likely be a longish, fact-based rant on the
subject.

- P

-- 
Peter N. M. Hansteen, member of the first RFC 1149 implementation team
http://bsdly.blogspot.com/ http://www.bsdly.net/ http://www.nuug.no/
"Remember to set the evil bit on all malicious network traffic"
delilah spamd[29949]: 85.152.224.147: disconnected after 42673 seconds.



Re: Routing 10-40 Mpps on OpenBSD

2016-09-11 Thread Chris Cappuccio
K [k...@protonmail.com] wrote:
> All,
> 
> This message is a call for people who are interested to benchmark commodity
> hardware with the goal of pushing as much PPS as possible through OpenBSD.
> The initial target is to reach 10 Mpps at 64 bytes (or more precisely 84
> bytes with interpacket gap) and if the experiment proves to be successful,
> we would then aim at 40+ Mpps.
> 
> The ultimate goal of this experiment is to build and share with the
> community a recognized hardware configuration that provides a good ground
> for real-world traffic at a typical small ISP.
> 
> We couldn't find such information online. In our case, the final setup
> would be two routers, each with two 10 Gbps uplink to upstreams Internet
> providers and an OSPF and iBGP connection between them. The software
> stack would be based on OpenBSD, OpenBGPD and OpenOSPFD. There is no
> commercial idea around the finding of this experiment.
> 
> While our budget is not unlimited and privately funded (by individuals),
> we are open to hear what hardware specifications people on this list
> would be interested to see. At the moment, we aim for this:
> 
> CPUs: Intel Xeon CPU E5-2697v2, E5-2667v2, E5-2680v3, E5-2640v3
> Intel NICs: Intel 82599ES, X520, X540-{T1/T2/AT2}, 85595, 82598,
> AF/82598, AT/82598, EB/82599, EB/82599 EN
> Chelsio NIcs: Chelsio T540-CR (although not sure there is an OpenBSD driver)

I think Intel and Myricom are going to be the best-supported 10GbE on
OpenBSD at the moment.

The best performance today will be with a processor that packs a lot
of punch into a smaller number of cores. I'm using Xeon E5-1630 v3
right now. The E5-2xxx series tend to have more cores at lower clock
speeds. They make more sense on a regular server. 

There is a lot of ongoing work in this area, OpenBSD doesn't claim to
be the performance leader today. 

Chris



Routing 10-40 Mpps on OpenBSD

2016-09-11 Thread K K
// Previous email bounced, so I resend it. Sorry for duplicate //

All,

This message is a call for people who are interested to benchmark commodity
hardware with the goal of pushing as much PPS as possible through OpenBSD.
The initial target is to reach 10 Mpps at 64 bytes (or more precisely 84
bytes with interpacket gap) and if the experiment proves to be successful,
we would then aim at 40+ Mpps.

The ultimate goal of this experiment is to build and share with the
community a recognized hardware configuration that provides a good ground
for real-world traffic at a typical small ISP.

We couldn't find such information online. In our case, the final setup
would be two routers, each with two 10 Gbps uplink to upstreams Internet
providers and an OSPF and iBGP connection between them. The software
stack would be based on OpenBSD, OpenBGPD and OpenOSPFD. There is no
commercial idea around the finding of this experiment.

While our budget is not unlimited and privately funded (by individuals),
we are open to hear what hardware specifications people on this list
would be interested to see. At the moment, we aim for this:

CPUs: Intel Xeon CPU E5-2697v2, E5-2667v2, E5-2680v3, E5-2640v3
Intel NICs: Intel 82599ES, X520, X540-{T1/T2/AT2}, 85595, 82598,
AF/82598, AT/82598, EB/82599, EB/82599 EN
Chelsio NIcs: Chelsio T540-CR (although not sure there is an OpenBSD driver)

If you consider other hardware options, please feel free to reply and let us
know.
We surely will not be testing all these configurations, we will most likely
pick on
CPU from the list and 2-3 NICs from the list as well. This experiment might be
also
taken to FreeBSD for comparison. If necessary, we consider sending this
configuration in a test center with Spirent hardware to validate this.

Feedbacks, questions, remarks, doubts, irony, are all welcome :-)

Cheers.



Routing 10-40 Mpps on OpenBSD

2016-09-11 Thread K
All,

This message is a call for people who are interested to benchmark commodity
hardware with the goal of pushing as much PPS as possible through OpenBSD.
The initial target is to reach 10 Mpps at 64 bytes (or more precisely 84
bytes with interpacket gap) and if the experiment proves to be successful,
we would then aim at 40+ Mpps.

The ultimate goal of this experiment is to build and share with the
community a recognized hardware configuration that provides a good ground
for real-world traffic at a typical small ISP.

We couldn't find such information online. In our case, the final setup
would be two routers, each with two 10 Gbps uplink to upstreams Internet
providers and an OSPF and iBGP connection between them. The software
stack would be based on OpenBSD, OpenBGPD and OpenOSPFD. There is no
commercial idea around the finding of this experiment.

While our budget is not unlimited and privately funded (by individuals),
we are open to hear what hardware specifications people on this list
would be interested to see. At the moment, we aim for this:

CPUs: Intel Xeon CPU E5-2697v2, E5-2667v2, E5-2680v3, E5-2640v3
Intel NICs: Intel 82599ES, X520, X540-{T1/T2/AT2}, 85595, 82598,
AF/82598, AT/82598, EB/82599, EB/82599 EN
Chelsio NIcs: Chelsio T540-CR (although not sure there is an OpenBSD driver)

If you consider other hardware options, please feel free to reply and let us 
know.
We surely will not be testing all these configurations, we will most likely 
pick on
CPU from the list and 2-3 NICs from the list as well. This experiment might be 
also
taken to FreeBSD for comparison. If necessary, we consider sending this
configuration in a test center with Spirent hardware to validate this.

Feedbacks, questions, remarks, doubts, irony, are all welcome :-)

Cheers.