Re: NAPI, e100, and system performance problem

[Greg Banks <gnb@sgi.com>]

On Fri, Apr 22, 2005 at 02:18:22PM -0400, jamal wrote:
> On Fri, 2005-22-04 at 19:21 +0200, Andi Kleen wrote:
> > On Fri, Apr 22, 2005 at 08:33:15AM -0400, jamal wrote:
> [..]
> > > They should not run slower - but they may consume more CPU.
> > 
> > They actually run slower.
> > 

IIRC I saw a similar but very small effect on Altix hardware about 18
months ago, but I'm unable to get at my old logbooks right now.  I
do remember the effect was very small compared to the CPU usage effect
and I didn't bother investigating or mentioning it.

> Why do they run slower? There could be 1000 other variables involved?
> What is it that makes you so sure it is NAPI?

At the time I was running 2 kernels identical except that one had
NAPI disabled in tg3.c.

> There is only one complaint I have ever heard about NAPI and it is about
> low rates: It consumes more CPU at very low rates. Very low rates
> depends on how fast your CPU can process at any given time. Refer to my
> earlier email. Are you saying low rates are a common load?
> 
> The choices are: a) at high rates you die or b) at _very low_ rates
> you consume more CPU (3-6% more depending on your system). 

This is a false dichotomy.  The mechanism could instead dynamically
adjust to the actual network load.  For example dev->weight could
be dynamically adjusted according to a 1-second average packet
arrival rate on that device.  As a further example the driver could
use that value as a guide to control interrupt coalescing parameters.

In SGI's fileserving group we commonly see two very different traffic
patterns, both of which must work efficiently without manual tuning.

1.  high-bandwidth, CPU-sensitive: NFS and CIFS data and metadata
    traffic.

2.  low bandwidth, latency-sensitive: metadata traffic on SGI's
    proprietary clustered filesystem.

The solution on Irix was a dynamic feedback mechanism in the driver
to control the interrupt coalescing parameters, so the driver
adjusts to the predominant traffic.

I think this is a generic problem that other people face too, possibly
without being aware of it.  Given that NAPI seeks to be a generic
solution to device interrupt control, and given that it spreads
responsibility between the driver and the device layer, I think
there is room to improve NAPI to cater for various workloads without
implementing enormously complicated control mechanisms in each driver.

> Logic says lets choose a). You could overcome b) by turning on
> mitigation at the expense of latency. We could "fix" at a cost of 
> making the whole state machine complex - which would be defeating  
> the " optimize for the common".

Sure, NAPI is simple.  Current experience on Altix is that
NAPI is the solution that is clear, simple, and wrong.

> >> Note, this would entirely solve what Andi and the SGI people are 
> >> talking about.
> > 
> > Perhaps, but Linux has to perform well on old hardware too.
> > New silicon is not a solution.

Agreed.

Greg.
-- 
Greg Banks, R&D Software Engineer, SGI Australian Software Group.
I don't speak for SGI.