Peter Memishian wrote:
>  > > Would the properties still end up in the show-linkprop output?  If so,
>  > > even well-behaved admins are still subjected to a mountain of properties
>  > > that they shouldn't even be using, which seems bad.  Or are you proposing
>  > > some way to hide these properties?  (I'm nervous about that as well.)  If
>  > > we're sure this stuff will be off the beaten path, maybe we could reduce
>  > > the number of properties to just `cap', `advcap', and `lpadvcap' (or
>  > > somesuch) with a commas-separated list of values?
>  > 
>  > What we really need is a simple UI for managing the "common" tunables 
>  > (and maybe exposing them as "properties" isn't the best way to handle 
>  > this!  see for example the way we handle WIFI "properties" for SSID... 
>  > they have their own subcommands)
>
> Right, that was the thinking behind that breakdown.  But I think most of
> the tunables we've been discussing *aren't* that common -- at least, I was
> kind of hoping most customers aren't mucking with fixed speed and duplex
>   

I think that hope is in vain.  I still hear a lot of sites that have a 
stupid policy of forcing everything to some mode in their network 
(usually forcing everything to 100 full duplex.)   We can  debate the 
wisdom of this (or lack thereof, in which I suspect we are in 
agreement), but the reality is our customers are *still* demanding this 
support.

> settings.  Jumbograms are in a bit of gray area, but if we had a boolean
> jumbograms property as you suggest, I think that'd be adminstratively
> decent.

One thing that we don't have is a nice way to separate out these "gray 
areas" (or areas where we expect customer tuning to be reasonable) from 
the areas where we expect tuning to be reasonable.  I would prefer, all 
things being equal, if customers who didn't care about 
rx_bcopy_threshold settings were not exposed to them.  (This tunable may 
need to be customer accessible, but it's reasonable that this adjustment 
require popping the hood of the automobile rather than being associated 
with a knob on the steering column.)

The jumbo-frame, and also the link speed/duplex setting, by comparison, 
are the sort of thing that we need to be reasonably accessible to 
minimize service calls.  I think of them as being like the traction 
control or over-drive adjustments in many late model cars... readily 
available for the folks who need them -- without checking the owner's 
manual or popping the hood, even though most people will probably never 
need to touch either one of them.  I guess you could also liken them to 
the first or second gear selections on a vehicle with an automatic 
transmission as well. :-)


>   I'm curious why you think having three properties (ala cap,
> advcap and lpadvcap) wouldn't be a reasonable tradeoff?
>   

Because:

a) that doesn't tell the whole story.  (is autonegotiation enabled?  
that's another boolean!  and having it disabled is a distinct case from 
having it enabled with only one mode advertised.)

b) i guess having adv_cap be tunable via a comma delimited list would 
work, but it is likely to be more confusing than just "please force this 
mode" (which is what admins who diddle here really want)

c) just setting a single forced mode is all (arbitrary unsubstantiated 
statistic alert!) 99% of the customers who do diddle here need...

Given that this is the state of affairs, lets make case "C" above as 
easy as possible.

>  > Actually, the plethora of different jumbo frame sizes that various NICs
>  > support is *another* annoyance.  I think we should just pick a standard
>  > Solaris value that is typical... say 9000 bytes or somesuch, and then
>  > abandon the attempts to support other less common "jumbo sizes".  The
>  > infinite tunability here is more likely to cause problems for customers
>  > than to help them.)
>
> I'm inclined to agree, but I'm not familiar with what causes a customer to
> choose a given value.
>
>   
I believe that most gigabit equipment supports 9000.  Some equipment 
supports larger sizes, some (very uncommon) support some size > 1518 but 
less than 9000.

I think its time someone here took a stand.   9000 bytes is a good size 
because it is nearly universal, and is sufficient to hold an 8K NFS block.

A very few devices support 16K or even higher, but the utility of such 
large MTUs is probably somewhat limited... even at 10Gb.  And the 
interoperability concerns with such large MTUs abound.  Further, the 
ethernet CRC-32 breaks down somewhere around 12K, so larger frames than 
12K are a bad idea with ethernet. 

A few references:

http://sd.wareonearth.com/~phil/jumbo.html
http://www.networkworld.com/forum/0223jumboyes.html
http://en.wikipedia.org/wiki/Super_jumbo_frame
http://www.sun.com/products/networking/ethernet/jumbo/index.html

That last reference suggests that we (Sun) support 9K frames, which 
while not entirely uncommon, is nothing near as common as the normally 
accepted value of 9000 bytes.

    -- Garrett

Reply via email to