Steve,
I’m sorry to get back to this so late. We do have new spam control software,
that
treats every outside msg as a spam. Sounds like I’m back to IBM in the 80s
:-).
Yes it is a very interesting thought, and might very well turn out to be a
vaperware !
I did not look into the diffserv in detail. My main goal is to findout what
MS provides.
As I mentioned about a month ago, that we have every reason to believe that
TCP stack is a potential bottleneck.
Today, I found some interesting detail(s) about MS’s implementations of the
whole traffic mgmt business. There is a guy
from MS wrote the book, and it is only for 2k. XP+ avoids RSVP signalling,
as we know RSVP had bad lucks from the
inception in 1994, and the control-signalling is not scaleable. That was
ideal for establishing virtual circuit on shared media (CSMA/CD(CA) I did
have a cursory look at cisco’s admission controls, but as I said there are a
wide variety of routers ( upto layer 3 protocols) and they might very well
ignore some control-signal implementation(s).
One of the criteria is to have MS workstation oses (2k, Xp) where my
playfield would be, so any server OS is out of the question(s), no domain
controller, no LDAP those are too shavvy.
On XP quite a bit of SBM, ACS and others are taken out, and it might be that
apps needs to use GQoS, but my point is totally heterogenous vendors, after
all its a free world sometime 
At least I’m now able to look at the main concerns decisively, before I was
bit confused. Lets see !. Coming up with different pacing/shaping algorithms
would not be a problem, it is to figureout where the defusion has to occur.
Once I get some interesting approach, I will let you know.
Thanks very much for your thoughtful reply.
-prokash
-----Original Message-----
From: xxxxx@lists.osr.com
[mailto:xxxxx@lists.osr.com]On Behalf Of Steve Dispensa
Sent: Wednesday, March 10, 2004 7:55 AM
To: Windows System Software Devs Interest List
Subject: Re: [ntdev] Admission Control
On Tue, 2004-03-09 at 22:21, Prokash Sinha wrote:
For simplicity, assume just a home-network, where one or more router with
some machines can act as high-intensity
server(s), and rest of them are clients, clients can be from PDA to
full-fldge server(s), and they could be from different vendors. WE ONLY
HAVE
CONTROLS OVER THOSE HIGH-INTENSITY SERVER(S) and they run on WIn2k/Xp
Home/Pro.
We dont want to assume that the clients are QOS enabled, only thing we
assume is the routers might have layer2 and layer 3 bandwidth mgmt
controls
( for example rsvp, may be not a good example but one of the others could
be
a possibility).
If you’re using IntServ or something in the network core, you don’t
strictly need any support for admission control, even in the servers, do
you?
When a request comes to one of our server, we would like to see if the
network bandwidth permits such a requested load.
Are you talking about aggregate network bandwidth or end-to-end
bandwidth between the server and the requesting client?
It is possible that there
might be more than one route between the src and dest, but for simplicity,
just take one route. RSVP seems to be an alternative, but XP onwards are
not
supporting directly, even though GQOS might have this under the hood along
side with other thing(s). Again for simplicity take some avg intensity (
no
poisson estimate to have exponential inter arrival, pls dont get into
this ), also ignore burst and little payload, just assume some constant
avg
rate).
So you’re looking for a first-order measurement of some sort? Or maybe
something like a smoothed average ( n * c + (1-n) * p, n<1 = weight on
current sample (c), p = previous sample), which is pretty easy to
calculate on the fly.
Question is what API’s I should look at!. Is TC Api good enough. Assuming
to
control any traffic comes thru winsock/afd. Is layered service provider an
alternative?. BASICALLY THE IDEA IS TO TRAVERSE THRU THE ROUTE FROM
PRODUCER
TO CONSUMER AT ANY GIVEN TIME, AND ASK ANYONE ENROUTE IF IT CAN HANDLE
SOME
PAYLOAD, IF SO RESERVE IT (RSVP seems like a good fit !).
Ah, yeah, that’s different. You could implement RSVP easily enough
without messing with the hosts, though; check out
http://www.cisco.com/univercd/home/home.htm - all of Cisco’s technical
documentation online.
Like to have as minimally intrusive as possible !!!
Network-based QoS seems like a good fit, then. LSPs are not (see
another post on the same topic). IM drivers are OK, but if you want
backward compatibility to 98 and PDA support, they’re out too. You may
well need different solutions on each platform.
Unrelated, have you looked into DiffServ? That is often sufficient for
bandwidth management needs, and it’s stateless to boot, so it’s more
efficient on the network.
What is the nature of the problem you’re trying to solve? Is it
reasonable to solve it in network hardware only?
Sounds like an iinterestnig project!
-sd
Questions? First check the Kernel Driver FAQ at
http://www.osronline.com/article.cfm?id=256
You are currently subscribed to ntdev as: xxxxx@garlic.com
To unsubscribe send a blank email to xxxxx@lists.osr.com