G'day,

I have a (relatively) simple question to which I can't find the answer
online.

How many servers (assuming similar spec) can a single load balancer handle?

For example, could 1 pound server manage requests to 12 servers (I'm
assuming yes), what about 200 servers, or 2'000 servers.

I'm working on an idea around scaling web applications by making each node
on the grid a combination of an application server and a load balancer (so
it would offload requests it can't deal with to other servers). But this
ratio may be the critical number... :)

Any advice / suggestions you can offer would be greatly appreciated.

Cheers,


Callum.
http://www.callum-macdonald.com/

Anyone got any rough estimates on this? Or any performance data for how much
traffic a server can handle?

Cheers - Callum.

On 21/09/2007, Callum Macdonald
<callum.macdonald+pound-list(at)gmail.com>
wrote:[...]

Hello Callum,

I doubt you will get any meaningful answer. If you asked how many 
requests per second a single pound balancer can handle, that we can tell 
you (my rough estimate would be about 1000 req/sec per one processor, 
assuming you don't run out of IP resources). But how many backends? 
You'll have to try.

I see a design problem with your suggestion. This just isn't what pound 
was made for. How will you determine, that the local server cannot 
handle the request?

The best thing you can do is estimate how many balancers you will need 
and dedicate that many servers for balancing. Combining backend and 
balancer on one machine does not make much sense to me.

Bye,
Michal



Callum Macdonald napsal(a):[...][...]

[...]

Hi Michal,

Thanks for your response, that's exactly the sort of number I need.

The idea is not to use pound, but perhaps to develop a modified version of
pound.

My concept is that a group of machines will act as a "cloud" (nothing new
there). Each machine will connect to others on the network and exchange live
data about free CPU cycles and free memory. So in principle, all load
balancers will know the state of most of the other machines at any given
time. They will then allocate the traffic based on various algorithms.

The idea is that any machine on the network can receive the request, but
will offload the request to another machine on the grid if it needs to.

Potentially, if using something like Amazon EC2, additional machines could
be brought up automatically to cope with additional load. Then as load
reduces, machines could be taken offline automatically.

At this stage, it's just a concept I'm throwing around.

Thanks again for your feedback.

Cheers - Callum.

On 24/09/2007, Michal Taborsky - Internet Mall
<michal.taborsky(at)mall.cz>
wrote:[...]

Hmm, if it works, it will be interesting. We are running a more simple 
version of this. We have a cloud (read "two", but can be more) of load 
balancers and a cloud of application servers. We can add and remove 
servers from/to these clouds without interuption of service. But to put 
it all on one server, that's a new one. The question is, is the overhead 
worth it? Because I imagine the communication between the nodes will put 
some stress on the machines and as the number of nodes grows, it might 
spend more processor time (and electricity) doing the inter-node 
communication than actually doing any work.

If you get it up, let us know.

MT.

Callum Macdonald napsal(a):[...][...]
>>> Anyone got any rough estimates on this? Or any performance data
for how[...]
>>> traffic a server can handle?
>>>
>>> Cheers - Callum.
>>>
>>> On 21/09/2007, Callum Macdonald
<callum.macdonald+pound-list(at)gmail.com>
>>> wrote:
>>>> G'day,
>>>>
>>>> I have a (relatively) simple question to which I can't find
the answer
>>>> online.
>>>>
>>>> How many servers (assuming similar spec) can a single load
balancer
>>>> handle?
>>>>
>>>> For example, could 1 pound server manage requests to 12
servers (I'm
>>>> assuming yes), what about 200 servers, or 2'000 servers.
>>>>
>>>> I'm working on an idea around scaling web applications by
making each[...]
>>>> on the grid a combination of an application server and a load
balancer[...]
>>>> it would offload requests it can't deal with to other
servers). But[...]
>>>> ratio may be the critical number... :)
>>>>
>>>> Any advice / suggestions you can offer would be greatly
appreciated.
>>>>
>>>> Cheers,
[...]

Hey Michal,

I'm only working on the concept at the moment. The potential practical
application is some months, probably years off. Once I get it into a
workable idea I'll gladly post a message back on the list though.

The idea behind it is that the cloud can be totally flexible. Currently,
you're always limited by your primary load balancer, which is also a single
point of failure. My model could be deployed behind a dynamic round robin
dns system for example. In principle the traffic should spread evenly over
the nodes, but if not, the system can deal with it.

I also like the peer to peer aspect, whereby nodes can be brought online and
dynamically join the cloud and pick up a share of the load.

I'll keep the list posted.

Cheers - Callum.

On 25/09/2007, Michal Taborsky - Internet Mall
<michal.taborsky(at)mall.cz>
wrote:[...]