At Tue, 5 Jun 2007 15:28:25 -0400, David Maxwell wrote:
Subject: Re: Looking ahead
>
> Consider the case of hot-swap interface cards in a router...
If it's a really well designed router architecture then you may not even
have actual drivers in the kernel for each type of card that can be
hot-swapped into the box -- i.e. the actual unix-y kernel will only be
supervisory and the protocols used to do admin tasks with each
hot-swappable hardware device will be standardized (e.g. as with Juniper).
Also I think if you want to support true hot-swap capabilities with some
form of loadable device driver support then the current LKM
infrastructure is completely unusable and would need a fresh re-design
anyway.
Obviously there are all kinds of ways to design such things and all
kinds of reasons to want to be able to load different drivers and other
kernel code into a running system.
But still there a whole lot of crazy issues to worry about when you
start considering the possibilities of modifying running code,
especially when you're modifying the running hardware underfoot at the
same time.
These days I think I'd start out looking at virtualization and migration
if at all possible, rather than trying to design brute-force HA from the
ground up in hardware.
Heck when you can get things like a single chip that does 20 10gigE
ports with full VLAN support, you can do some pretty wonky stuff just by
moving the outside wires and some processes and/or virtual machines
around and all without having to support true hot-swap in the very most
complex core hardware and kernel of a general purpose computing platform.
--
Greg A. Woods
H:+1 416 218-0098 W:+1 416 489-5852 x122 VE3TCP RoboHack
<woods%robohack.ca@localhost>
Planix, Inc. <woods%planix.com@localhost> Secrets of the Weird
<woods%weird.com@localhost>
Attachment:
pgpSjhXIW29rD.pgp
Description: PGP signature