OK, I'm still working on the new scheduler design, and I'd like the
development to be as open as possible for several reasons:
* To make sure everyone feels ownership, and has a say in the
development
* To take advantage of all the smart people out there
However, I have several times drafted e-mails for two aspects I'm
working on now (credit handling and load balancing), and ended up
waiting for several reasons. First, to understand where I was, I
thought it necessary to summarize the path that brought me there:
i.e., why not use the existing system, or use it with minor tweaks.
This always ends up taking a long time, and in the end it's hard to
explain completely for the same reason that I'm still a little stuck:
I don't understand it yet fully myself. Secondly, load balancing and
credits are actually rather connected: one cannot be adequately solved
without the other. Finally, there's always a disconnect between what
one might expect to happen with a certain algorithm, and what actually
happens in a given workload. So going into great lengths about the
design (and possibly involving everyone's expertise), only to
re-design it again after testing, I feared would risk wearying
people's patience.
So, I'm thinking about proceeding this way.
First, everyone should understand that currently things are in
development. So if getting involved in tossing out ideas and
experimentation in this area interests you, read the mail. Otherwise,
feel free to ignore it until I present it as a more finalized design.
In order to facilitate this, I'm going to change the subject line from
[RFC] to [ED] (Experimental development). When things seem more
solid, and I'm presenting something that I want more developers to
actually take a look at and check, I'll change the subject line back
to [RFC].
Secondly, I'm going to describe my ideas for load balancing in an
e-mail, and then throw together a basic version with a patch that
people can download and play around with. That should get us close
enough to my load-balancing vision to be able to play with credits.
I'm keen, once I get some basic prototypes working, to see how it runs
for a wide number of workloads. So I'd appreciate anyone who's
interested.
Any feedback / advice on this plan is welcome.
Peace,
-George
_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel
|
Home