Default offload settings in Ethernet drivers

Default offload settings in Ethernet drivers

[Jan Ceuleers]

Hi!

A discussion recently took place on the power mailing list on the 
subject of the impact of (hardware-assisted) offload functions on the 
power efficiency of the overall system.

The discussion was brought on by me noticing that not all drivers enable 
  all of their offload features by default (case in point: r8169).

Although the discussion may not be complete, early indications are that:

1. Hardware-assisted offloads improve power efficiency unless 
implemented in a separate CPU (TOE / Total Offloading);

2. It would probably be a good idea to enable hardware-assisted offloads 
other than TOE by default given the above.

I would therefore like to sollicit views here:

1. Would changing default offload settings in Ethernet drivers help to 
save the planet?

2. Which offload settings does it make sense to enable by default?


Thanks, Jan

Re: Default offload settings in Ethernet drivers

[Stephen Hemminger]

On Thu, 11 Dec 2008 19:41:31 +0100
Jan Ceuleers <jan.ceuleers@computer.org> wrote:

> Hi!
> 
> A discussion recently took place on the power mailing list on the 
> subject of the impact of (hardware-assisted) offload functions on the 
> power efficiency of the overall system.
> 
> The discussion was brought on by me noticing that not all drivers enable 
>   all of their offload features by default (case in point: r8169).
> 
> Although the discussion may not be complete, early indications are that:
> 
> 1. Hardware-assisted offloads improve power efficiency unless 
> implemented in a separate CPU (TOE / Total Offloading);
> 
> 2. It would probably be a good idea to enable hardware-assisted offloads 
> other than TOE by default given the above.
> 
> I would therefore like to sollicit views here:
> 
> 1. Would changing default offload settings in Ethernet drivers help to 
> save the planet?
> 
> 2. Which offload settings does it make sense to enable by default?

Go get a kill-a-watt meter and real hardware and measure.
I don't think there will be any difference. Linux (mainline) doesn't
do TOE. It does do segmentation offload, and any driver that can
do segmentation offload enables it by default.

Re: Default offload settings in Ethernet drivers

[Rick Jones]

> Go get a kill-a-watt meter and real hardware and measure.
> I don't think there will be any difference. Linux (mainline) doesn't
> do TOE. It does do segmentation offload, and any driver that can
> do segmentation offload enables it by default.

Further, it would seem reasonable to expect if a NIC has a feature the 
driver does not enable by default, the implication is that feature is 
not yet ready for regular use.

rick jones

Re: Default offload settings in Ethernet drivers

[Ben Hutchings]

On Thu, 2008-12-11 at 10:54 -0800, Stephen Hemminger wrote:
> On Thu, 11 Dec 2008 19:41:31 +0100
> Jan Ceuleers <jan.ceuleers@computer.org> wrote:
> 
> > Hi!
> > 
> > A discussion recently took place on the power mailing list on the 
> > subject of the impact of (hardware-assisted) offload functions on the 
> > power efficiency of the overall system.
> > 
> > The discussion was brought on by me noticing that not all drivers enable 
> >   all of their offload features by default (case in point: r8169).
> > 
> > Although the discussion may not be complete, early indications are that:
> > 
> > 1. Hardware-assisted offloads improve power efficiency unless 
> > implemented in a separate CPU (TOE / Total Offloading);
> > 
> > 2. It would probably be a good idea to enable hardware-assisted offloads 
> > other than TOE by default given the above.
> > 
> > I would therefore like to sollicit views here:
> > 
> > 1. Would changing default offload settings in Ethernet drivers help to 
> > save the planet?
> > 
> > 2. Which offload settings does it make sense to enable by default?
> 
> Go get a kill-a-watt meter and real hardware and measure.
[...]

Even then, the results will be highly dependent on the CPU's power-
saving capabilities and on settings that affect the pattern of IRQs like
interrupt moderation and number of queues used by multiqueue-capable
drivers, not just on the offload settings.  I would expect checksum
generation/validation and segmentation in an ASIC to take less power
than in a CPU, but on an already-busy CPU this might not be the case.

Power usage also depends on throughput, of course.  If the test involves
pushing data as fast as possible rather than simulating a specific
workload then offload features may well probably increase throughput
without reducing power consumption.  So maybe the metric should be
power/throughput... but there is unlikely to be a linear relationship
between the two, so a single figure for this may be misleading.

Ben.

-- 
Ben Hutchings, Senior Software Engineer, Solarflare Communications
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.

Re: Default offload settings in Ethernet drivers

[Jan Ceuleers]

Stephen Hemminger wrote:
> I don't think there will be any difference. Linux (mainline) doesn't
> do TOE. It does do segmentation offload, and any driver that can
> do segmentation offload enables it by default.

Stephen,

Thanks for your reply. My experience with the r8169 driver is that it 
does support rx checksumming, tx checksumming, scatter/gather and tcp 
segmentation, but out of the above only enables rx checksumming by default.

The other two ethernet drivers I have hardware for don't support any 
hardware-assisted offloads (that is: they all support generic 
segmentation offload but that isn't hardware-assisted and it is also not 
enabled by default).

Best regards, Jan

Re: Default offload settings in Ethernet drivers

[Ben Hutchings]

On Thu, 2008-12-11 at 20:30 +0100, Jan Ceuleers wrote:
> Stephen Hemminger wrote:
> > I don't think there will be any difference. Linux (mainline) doesn't
> > do TOE. It does do segmentation offload, and any driver that can
> > do segmentation offload enables it by default.
> 
> Stephen,
> 
> Thanks for your reply. My experience with the r8169 driver is that it 
> does support rx checksumming, tx checksumming, scatter/gather and tcp 
> segmentation, but out of the above only enables rx checksumming by default.
> 
> The other two ethernet drivers I have hardware for don't support any 
> hardware-assisted offloads (that is: they all support generic 
> segmentation offload but that isn't hardware-assisted and it is also not 
> enabled by default).

Since v2.6.27, GSO is enabled by default for devices that can do TX
scatter/gather and checksum offload.

Ben.

-- 
Ben Hutchings, Senior Software Engineer, Solarflare Communications
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.