|
Hello jamal,
On Mon, 15 May 2000, jamal wrote:
>
>
> On Mon, 15 May 2000, Julian Anastasov wrote:
>
> >
> > Yes. This is only one real example. With keep alive
> > ON, most of the connections can be closed first from the
> > client.
>
> But it does explain the reduced ratio you see ;->
Yes, sorry :)
> > > 2.3 networking scales linearly with number of processors. Example typical
> > > routing max capacities are around 30-40K packets/sec; same in 2.3 but if
> > > you add another procesor, you end somewhere in the 60-70Kpps with 2.3 (2.2
> > > doesnt change).
> >
> > You forgot to describe the hardware :)
>
> Sorry, this was the same motherboard with one processor and then two
> processors so the numbers are relative; a really shitty motherboard known
> as the ABIT6 which can take upto 2 400Mhz (?)celeron processors.
> 64MB of RAM.
OK
> > > your connection setup because it includes a little more overhead than
> > > typical routing should get less performance.
> >
> > Right. With faster enough CPU this is only latency,
> > not less performance! If we start to drop packets this is
> > bad.
>
>
> I would think it would be a little more complicated than that.
> Depends on a lot of other tunings (simple example: if you use DMAing NICs,
> how big are those receive Ring/lists). Latency would probably hover around
> the PCI latency then.
I agree. There are so many things involved. May be
we will receive help from the gurus about tuning the LVS
performance on SMP: which primitives to prefer, etc.
> > Lets see his results :)
>
> I hope he's going to do it.
> >
> > Well, I didn't performed such tests. That doesn't
> > mean that there are no such tests. Please, I'm not the only
> > LVS user :)
> >
>
> I thought you were one of the developers (in which case this should
> concern you).
Yes, but I don't have resources for such tests, at
least on such high-end level. There are other people who
performed some tests. I can only help the LVS code to work
better. Of course, I'm interested too see such tests and
their results. In these days when the hardware speed
changes so rapidly we can be always surprised for the
numbers.
> > I don't know your needs, really! If this hardware
> > is faster and cheaper and you think that is enough just save
> > your money and don't buy PC :) The people have different
> > needs. One needs PIII 2GB LVS box another can live with
> > 128MB box :) I can't give you the universal solution.
> >
> > The only thing I can do is to explain the details.
> > The users always have questions about the LVS internals. I
> > can't tell you "You must use LVS - It is free". There are
> > so many things that must be considered. May be the simpler
> > answer is "Buy hardware for load balancing. It has good TCP
> > stack and it is without bugs and ...". May be this
> > hardware is better, may be the support is better. Why not!
> > Why the only load balancing solution must be LVS?
>
> I am just a curious observer as far as LVS is concerned. I have absolutely
> no need for it (maybe at some later point, since i know exists).
>
> I am not asking you to justify why LVS exists or why people should use
> LVS; i am just making a statement that load balancing h/ware is becoming
> commodity. If the trend continues, you should be able to buy a black box
> worth $200 which will have the same features as LVS does; perhaps i am
And may be other features too :)
> exagerating a bit, but if this is the case, why should someone continue to
> use LVS? I was hoping you will enlighten me in some way instead you become
> defensive and give me a marketing pitch.
I will try but you know everything that I will say.
Be more specific. Your question "why should someone
continue to use LVS?" is not correct. This is like "How LVS
will save the world" :) Or "Why should I buy this hardware
when I can use software (LVS) ?". What can I say. Because
it is fast? How faster, you ask :) I can't compare LVS with
something I don't know. And $200 is the only we know. How
you will put 512MB RAM in $200 box :) Don't surpise us with
such facts until we have something to compare :) What about
if I say "LVS will cost you $1" because the director box can
be used also as web server. You buy 2 PC for web and run
LVS on one of them :) We just saved $200 by not buying
hardware!!! Or may be this hardware can run web?
And the LVS features are present in the web site.
What is the information that is not visible? The internals?
I can give you answers but for specific questions :)
> Maybe someone else has an answer.
Yes, may be someone not involved in our thread.
I agree with you!
But after giving this price of $200 we have to pay
to someone to maintain this box. May be the price of the
product is not meaningful! Everyone must evaluate the costs
for maintainance. If we compare the product price and the
costs after buying the product, for example, for a high-end
cluster, the first one will be 1/n of the other. Sorry for
my english but you understand me. So, the initial price is
may be not meaningful. The rest is the performance and the
maintainance. Here we give some facts with you but the
whole picture is not in just buying the balancer. We have
different users with different abilities to maintain such
product. They will decide which is the better variant for
them. Of course, may be LVS don't have some features that
are present in the other balancers. We accept many ideas
for improvements but we try to keep the code simpler and to
implement features which are useful for many users.
We are very happy that in linux 2.[34] we can
increase the director throughput. We will try to use the
per-processor networking in the new kernels and to improve
the access to the shared data.
You agree, our dispute is only theoretical and
hypothetical. I can't give you the numbers you want. They
change. But examples exist on the web site. Sometimes they
appear in the mailing list too. Isn't that a marketing
pitch :)
Regards
--
Julian Anastasov <uli@xxxxxxxxxxxxxxxxxxxxxx>
|
