On (03/04/09 15:32), Darren Reed wrote:
> On  4/03/09 05:59 AM, Sowmini.Varadhan at Sun.COM wrote:
>> ...
>>
>> Again, the question is, "who is using this?". For far
>> too long we have speculated that a tunable "might be useful"
>> and carried this baggage around, without actually checking
>> if the tunable is actually used.
>>   
>
> Please explain to me how having some value that is used in IP
> is worse off for being used via a variable (that can be tuned.)

Note that I am not saying that we should not have any tunables.
But we have a long list of "just in case" tunables, and we need
to re-inspect what that "case" was, that they were designed for. 

We can make everything a variable. Then the user is left
with the question "what do you want me to do with this?"
and we are left wondering "why is this customers performance
different from the rest?"

Do you really enjoy, for example, figuring out what softrings
parameters to use for e1000g, or what parameters "work best"
for nxge? Or figuring out when to set ip_output_queue?

To quote Jim:
"..It feels like a manual choke on a gasoline engine or a 
 fine-tune adjust knob on a TV; when did you last have to mess with one
 of those?"

> Is it because you have address it with brussels?
> Is it because someone might have to document it?
> Is it because it makes the system measurably slower?
> Is it because it restricts the ways in which we design and
> architect IP in the future?

It is because we have a complex system with lots of knobs
that is hard to extend and debug. 

> Exactly what is this "baggage"?

that is what I am trying to figure out. I really dont think,
for examle, that "ip_multirt_ttl" is very useful. But maybe
you know some application that benefits from this? (If you
do, then yes, we should keep this).

--Sowmini


Reply via email to