Skip to contentSkip to conversation

network namespace ipv6 perfs

DISCUSSION Daniel Lezcano 5 participants 10 messages Mar 3, 2008 – Mar 5, 2008

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

Hi,

Some performance tests was made by Benjamin to watch out the impact of the network namespace. The good news is there is no impact when used with or without namespaces. That has been checked using a real network device inside a network namespace.

These results are consistent with the ones previously made for ipv4.

http://lxc.sourceforge.net/network/bench_ipv6_graph.php

Thanks to Benjamin who did all the performance tests :)

Regards
-- Daniel

Sauf indication contraire ci-dessus:
Compagnie IBM France
Siège Social : Tour Descartes, 2, avenue Gambetta, La Défense 5, 92400 Courbevoie
RCS Nanterre 552 118 465
Forme Sociale : S.A.S.
Capital Social : 542.737.118 ?
SIREN/SIRET : 552 118 465 02430

#1 lore raw
Quoted 9 lines · Daniel Lezcano

Hi,

Some performance tests was made by Benjamin to watch out the impact of the network namespace. The good news is there is no impact when used with or without namespaces. That has been checked using a real network device inside a network namespace.

These results are consistent with the ones previously made for ipv4.

http://lxc.sourceforge.net/network/bench_ipv6_graph.php

Thanks to Benjamin who did all the performance tests :)

In these results, may be, there is one thing that should be explained. It is the CPU utilization overhead in the 'veth' case.

Compared to physical devices or macvlan, veth interfaces don't benefit from hardware offloading mechanisms: i.e. checksums have to be computed by the soft. That explains the big overhead in CPU utilization when using this kind of virtual interface.

Benjamin

Quoted 67 lines

Regards
-- Daniel

Sauf indication contraire ci-dessus:
Compagnie IBM France
Siège Social : Tour Descartes, 2, avenue Gambetta, La Défense 5, 92400 Courbevoie
RCS Nanterre 552 118 465
Forme Sociale : S.A.S.
Capital Social : 542.737.118 ?
SIREN/SIRET : 552 118 465 02430

Signature · 3 lines
B e n j a m i n   T h e r y  - BULL/DT/Open Software R&D

   http://www.bull.com
#2 lore raw
Quoted 12 lines · Benjamin Thery
Quoted 11 lines · Daniel Lezcano

Hi,

Some performance tests was made by Benjamin to watch out the impact of the network namespace. The good news is there is no impact when used with or without namespaces. That has been checked using a real network device inside a network namespace.

These results are consistent with the ones previously made for ipv4.

http://lxc.sourceforge.net/network/bench_ipv6_graph.php

Quoted 8 lines
Quoted 1 line

Thanks to Benjamin who did all the performance tests :)

In these results, may be, there is one thing that should be explained. It is the CPU utilization overhead in the 'veth' case.

Compared to physical devices or macvlan, veth interfaces don't benefit from hardware offloading mechanisms: i.e. checksums have to be computed by the soft. That explains the big overhead in CPU utilization when

You can tune the veth devices not to account checksum when unnecessary.

Quoted 72 lines

using this kind of virtual interface.

Benjamin

Quoted 66 lines

Regards
-- Daniel

Sauf indication contraire ci-dessus:
Compagnie IBM France
Siège Social : Tour Descartes, 2, avenue Gambetta, La Défense 5, 92400 Courbevoie
RCS Nanterre 552 118 465
Forme Sociale : S.A.S.
Capital Social : 542.737.118 ?
SIREN/SIRET : 552 118 465 02430

#3 lore raw
Quoted 15 lines · Pavel Emelyanov
Quoted 14 lines · Benjamin Thery
Quoted 12 lines · Daniel Lezcano

Hi,

Some performance tests was made by Benjamin to watch out the impact of the network namespace. The good news is there is no impact when used with or without namespaces. That has been checked using a real network device inside a network namespace.

These results are consistent with the ones previously made for ipv4.

http://lxc.sourceforge.net/network/bench_ipv6_graph.php

Thanks to Benjamin who did all the performance tests :)

Quoted 8 lines
Quoted 6 lines

In these results, may be, there is one thing that should be explained. It is the CPU utilization overhead in the 'veth' case.

Compared to physical devices or macvlan, veth interfaces don't benefit from hardware offloading mechanisms: i.e. checksums have to be computed by the soft. That explains the big overhead in CPU utilization when

You can tune the veth devices not to account checksum when unnecessary.

Oh. This is interesting.

You mean with ethtool -K rx/tx?
I will give it a try.

Benjamin

Quoted 15 lines
Quoted 9 lines

using this kind of virtual interface.

Benjamin

Quoted 3 lines

Regards
-- Daniel

--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#4 lore raw
Quoted 23 lines · Benjamin Thery
Quoted 22 lines · Pavel Emelyanov
Quoted 20 lines · Benjamin Thery
Quoted 7 lines · Daniel Lezcano

Hi,

Some performance tests was made by Benjamin to watch out the impact of the network namespace. The good news is there is no impact when used with or without namespaces. That has been checked using a real network device inside a network namespace.

Quoted 5 lines

These results are consistent with the ones previously made for ipv4.

http://lxc.sourceforge.net/network/bench_ipv6_graph.php

Thanks to Benjamin who did all the performance tests :)

In these results, may be, there is one thing that should be explained. It is the CPU utilization overhead in the 'veth' case.

Compared to physical devices or macvlan, veth interfaces don't benefit from hardware offloading mechanisms: i.e. checksums have to be computed by the soft. That explains the big overhead in CPU utilization when

Quoted 6 lines
Quoted 1 line

You can tune the veth devices not to account checksum when unnecessary.

Oh. This is interesting.

You mean with ethtool -K rx/tx?
I will give it a try.

Pavel,

I had no luck with "ethtool -K veth0 rx on tx on".
On my testbed, with these options TCP drops packets
(trying to establish a ssh connection between init and child namespace).

Then, I tested "ethtool -K veth0 rx on tx off". This time TCP (and netperf) work, but I see no difference in CPU load compared to the case without offloading.

Can I tune veth differently?

(BTW, I run netperf between a child namespace on host A and netserv on host B. The stream goes through the following interface:
veth1 on A -> veth0 on A -> eth1 on A -> ("real network") -> eth1 on B)

Benjamin

Quoted 17 lines
Quoted 15 lines
Quoted 9 lines

using this kind of virtual interface.

Benjamin

Quoted 3 lines

Regards
-- Daniel

--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

Signature · 3 lines
B e n j a m i n   T h e r y  - BULL/DT/Open Software R&D

   http://www.bull.com
#5 lore raw
Quoted 37 lines · Benjamin Thery
Quoted 23 lines · Benjamin Thery
Quoted 22 lines · Pavel Emelyanov
Quoted 20 lines · Benjamin Thery
Quoted 7 lines · Daniel Lezcano

Hi,

Some performance tests was made by Benjamin to watch out the impact of the network namespace. The good news is there is no impact when used with or without namespaces. That has been checked using a real network device inside a network namespace.

Quoted 5 lines

These results are consistent with the ones previously made for ipv4.

http://lxc.sourceforge.net/network/bench_ipv6_graph.php

Thanks to Benjamin who did all the performance tests :)

In these results, may be, there is one thing that should be explained. It is the CPU utilization overhead in the 'veth' case.

Compared to physical devices or macvlan, veth interfaces don't benefit from hardware offloading mechanisms: i.e. checksums have to be computed by the soft. That explains the big overhead in CPU utilization when

Quoted 5 lines
Quoted 1 line

You can tune the veth devices not to account checksum when unnecessary.

Oh. This is interesting.

You mean with ethtool -K rx/tx?
I will give it a try.

Pavel,

I had no luck with "ethtool -K veth0 rx on tx on".
On my testbed, with these options TCP drops packets
(trying to establish a ssh connection between init and child namespace).

Quoted 5 lines

Then, I tested "ethtool -K veth0 rx on tx off". This time TCP (and netperf) work, but I see no difference in CPU load compared to the case without offloading.

Can I tune veth differently?

Yup. You may try turn tso and sg on as well.

Quoted 21 lines

(BTW, I run netperf between a child namespace on host A and netserv on host B. The stream goes through the following interface:
veth1 on A -> veth0 on A -> eth1 on A -> ("real network") -> eth1 on B)

Benjamin

Quoted 13 lines
Quoted 12 lines
Quoted 7 lines

using this kind of virtual interface.

Benjamin

Quoted 3 lines

Regards
-- Daniel

--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#6 lore raw
In reply to Daniel Lezcano

One more thing about these results: the kernel.
The version used to run these tests was 2.6.25-rc1 from Dave Miller's net-2.6 tree.

(and I included results from a vanilla 2.6.23.16 as reference)

Benjamin

Quoted 76 lines · Daniel Lezcano

Hi,

Some performance tests was made by Benjamin to watch out the impact of the network namespace. The good news is there is no impact when used with or without namespaces. That has been checked using a real network device inside a network namespace.

These results are consistent with the ones previously made for ipv4.

http://lxc.sourceforge.net/network/bench_ipv6_graph.php

Thanks to Benjamin who did all the performance tests :)

Regards
-- Daniel

Sauf indication contraire ci-dessus:
Compagnie IBM France
Siège Social : Tour Descartes, 2, avenue Gambetta, La Défense 5, 92400 Courbevoie
RCS Nanterre 552 118 465
Forme Sociale : S.A.S.
Capital Social : 542.737.118 ?
SIREN/SIRET : 552 118 465 02430

Signature · 3 lines
B e n j a m i n   T h e r y  - BULL/DT/Open Software R&D

   http://www.bull.com
#7 lore raw
In reply to Daniel Lezcano
Quoted 3 lines · Daniel Lezcano

Hi,

Some performance tests was made by Benjamin to watch out the impact of the network namespace. The good news is there is no impact when used with or without namespaces. That has been checked using a real network device inside a network namespace.

The *_RR tests seem to show a drop in througput and corresponding increases in service demand - could that be because things like TSO et al cannot mask much of anything in the way of a path-length increase?

From the annotations, I'm ass-u-me-ing that NS was only used on the netperf side and not both netperf and netserver side?

happy benchmarking,

rick jones

Quoted 6 lines

These results are consistent with the ones previously made for ipv4.

http://lxc.sourceforge.net/network/bench_ipv6_graph.php

Thanks to Benjamin who did all the performance tests :)

#8 lore raw
Quoted 6 lines · Rick Jones
Quoted 3 lines · Daniel Lezcano

Hi,

Some performance tests was made by Benjamin to watch out the impact of the network namespace. The good news is there is no impact when used with or without namespaces. That has been checked using a real network device inside a network namespace.

The *_RR tests seem to show a drop in througput and corresponding increases in service demand - could that be because things like TSO et al cannot mask much of anything in the way of a path-length increase?

Hmm. In fact Benjamin took the 2.6.23.16 kernel where there were no network namespace code at all. So these differences between 2.6.23.16 and 2.6.25-rc1 does not show a performance degradation especially related to the network namespaces. The important point is the 2.6.25-rc1 without ipv6 netns and 2.6.25-rc1 with ipv6 netns code applied, I mean the second and the third line and we can point that the ipv6 netns code does not degrade performances for either throughput and service demand.

Quoted 2 lines

From the annotations, I'm ass-u-me-ing that NS was only used on the
netperf side and not both netperf and netserver side?

right :)

Quoted 1 line

happy benchmarking,

Thanks Rick.

-- Daniel

#9 lore raw
Quoted 9 lines · Daniel Lezcano
Quoted 6 lines · Rick Jones
Quoted 3 lines · Daniel Lezcano

Hi,

Some performance tests was made by Benjamin to watch out the impact of the network namespace. The good news is there is no impact when used with or without namespaces. That has been checked using a real network device inside a network namespace.

The *_RR tests seem to show a drop in througput and corresponding increases in service demand - could that be because things like TSO et al cannot mask much of anything in the way of a path-length increase?

Hmm. In fact Benjamin took the 2.6.23.16 kernel where there were no network namespace code at all. So these differences between 2.6.23.16 and 2.6.25-rc1 does not show a performance degradation especially related to the network namespaces. The important point is the 2.6.25-rc1 without ipv6 netns and 2.6.25-rc1 with ipv6 netns code applied, I mean the second and the third line and we can point that the ipv6 netns code does not degrade performances for either throughput and service demand.

As Daniel stated, we should not compare the first bar with the other ones directly. May be I should have arranged the chart differently and made it more clear that the first bar is "2.6.23 vanilla" and the second one is "2.6.25-rc1 vanilla". Many changes happened in the whole kernel between 2.6.23 and 2.6.24 so we can't compare the first two bars to tell if network namespace degraded performances (and only a small part of netns is in 2.6.24).
The way I presented the chart is a bit misleading. :)

What's interesting to compare in the charts is the 2nd, 3rd and 4th lines. It shows that on the exact same hardware (in the 4th case the physical interface is moved into the child namespace), with or without the patchset, using network namespace or not, performance is about the same.

Benjamin

Quoted 12 lines
Quoted 2 lines

From the annotations, I'm ass-u-me-ing that NS was only used on the
netperf side and not both netperf and netserver side?

right :)

Quoted 1 line

happy benchmarking,

Thanks Rick.

-- Daniel

Signature · 3 lines
B e n j a m i n   T h e r y  - BULL/DT/Open Software R&D

   http://www.bull.com