Hi Yi,

I believe I have already given a couple of good reasons why as many parameters 
as possible should be tunable. And I also think that the approach to ask for a 
detailed explanation and use case for each and every parameter is not the right 
approach to the problem. So far I have only asked to implement the easiest 
possible solution (suggesting to implement per-session/connection/target 
tunables later), so the effort to implement these changes should be minimal and 
I would expect zero maintenance effort - plus a couple more happy customers.

I will summarize again my key arguments:

* There is no strong reasoning behind the chosen parameters, the now hardcoded
   values are relatively arbitrary. We need experience from the field to let
   evolve the values appropriate for most environments and we will most probably
   need to work on recommendations for various environments (e.g. we already 
have
   two cases: Optimizing for a fast failure vs. for a maximally robust
   connection).

   In short: We know of at least two cases where the current parameters dont
   match.

* In real world scenarios, customers will find themselves in situations where
   they would need to tune a parameter "now" to work around bugs or other 
issues,
   to  stabilize an unstable system etc. etc.

   If we only have compile time values, there is absolutely no way out of such
   situations and I guarantee that customers *will* complain to support about
   this.

   In short: We need flexibility.

* It does not harm to make #defines kernel tunables

I do second the argument that before implementing per session/connection/target 
tunables (with iscsiadm commands to control them), we need to be sure that 
there 
is a valid demand, but on the other hand I must say that so far I have not 
heard 
one sound argument why not to make these compile time defaults kernel tunables 
in the first step.

If you insist, I will try to find scenarios where tuning each and every 
parameter, but you will probably know that it is very hard to foresee use cases 
and failure scenarios - so whatever we come up with now will most likely be 
incomplete. And, again, I do think that my arguments are very valid already.

Community? And other opinions?

Nils
_______________________________________________
storage-discuss mailing list
[email protected]
http://mail.opensolaris.org/mailman/listinfo/storage-discuss

Reply via email to