On Fri, 21 Jan 2000, David Lang wrote:
> > I'd doubt though that all 3-400 of his users will be sitting on the same
> > Ethernet segment (if they are, the speed of the gateway won't much matter,
> > there will be worse problems.) The latency issues of 300 people all
> > hitting that gateway at once will be noticable (and measurable.) Enough
> > so that if the internal topology isn't laid out well, it's worth putting
> > in and outbound interfaces on the box just to drive down collisions on
> > the internal interface. Unfortunately, in current stable Linux kernels,
> > that means an immediate degragation in performance.
>
> can you go into more detail on this statement? the aratecture that I
> assumed for this is 100Mb full duplex ethernet to a switch which then has
> the internal users on it, on the outside i only have 2 T-1 lines so even
> 10Mb ethernet should be reasonable.
Let's say there are 400 users and 150 of the are active. Put them all on
a switched full duplex network, then take that aggragated traffic and push
it at even a 100Mb/s interface. (Remember that in a business, most
communications are internal, a school dorm probably won't have the same
pattern... unless everyone's playing Quake) You'll collide on the
internal NIC pretty heavily. I'd be worried about buffering on the
outbound side too.
> > "It works." is significantly different from "It works well." which is
> > still different from "This is as good as it gets." I'd expect a liberal
> > arts college to put heavier use on streaming protocols and that's where
> > latency can bite the most.
>
> Agreed, in my case it falls into the "it works well" catagory, we just
> switched out the linux box for a Sun ultra 10 333MHz 256M ram running
> Raptor, performance has not changed notisably.
That probably says more about Raptor- assuming you've done the usual
Solaris tuning tricks. That or you're at saturation point for your
outbound link (or buffers) or you're colliding trying to get to the Sun.
U10's have a PCI bus though :(
> > The first night everyone's stuck inside because of (pick your favorite
> > natural weather phenoninom) it'd suck to see a meltdown.
> >
> > With business users, you can figure on a high of about 15% active
> > concurrent sockets for your users. 20% is about the highest I've seen it
> > yet. In a college dorm in Iowa during a snowstorm I'd expect it to be
> > higher.
>
> Depends on the company :-) So far we have decided that until bandwidth
> starts to cost us more then it is now we do not object to realaudio
> streams, combining this with the fact that the building is in a
> radio black hole so people cannot bring in radios, there are a surprising
> number of people attaching to realaudio "stations" all day long wearing
> headphones while they work.
I'd expect music majors to use the higher CODECs and consume more
bandwidth, but I'd still recommend that Paul actually take a look at the
traffic patterns now and project some growth onto that and weigh it
against his current external bandwidth allocation. I wouldn't be too
surprised to see full recording-quality audio streams soon, and that'll
take a lot more bandwidth than RA. Our windows don't let radio in, and
we've more than a couple thousand lusers internally and a good ammount of
external bandwidth, and I still don't like what RA over HTTP does to the
old proxy server we have. The new ones seem to be a lot better, but we
scaled the old one back when Gopher was popular and the new ones aren't 3
months old :)
Paul
-----------------------------------------------------------------------------
Paul D. Robertson "My statements in this message are personal opinions
[EMAIL PROTECTED] which may have no basis whatsoever in fact."
PSB#9280
-
[To unsubscribe, send mail to [EMAIL PROTECTED] with
"unsubscribe firewalls" in the body of the message.]
