Network load balancing problem

  • Thread starter Thread starter Naresh
  • Start date Start date
N

Naresh

Hi,

In our network, we have three windows 2000 advanced
servers. They are configured with network load balancing
with two network adapters in unicast mode. The problem is
load balancing of the application is not working evenly on
all the three servers. This problem is intermittent. most
of the times, all the client connections go to the server
which has been configured with priority(unique host Id) as
1 and the other two servers will be with no load at all.
due to this, one server is heavily loaded and the other
two are completely free.

I have been searching the microsoft site for this problem.
all the settings for my network and servers are as per the
documentation.
Can any one help me out?
 
Naresh,
NLB is a statistical load-balancer based on the client identity (e.g.,
IP address and/or port). What that means is that a large client population
is necessary in order to perceive the configured "balance". How many
clients are you sending the traffic from?
--
Thanks,
Karthic
Software Design Engineer
Windows Server Group
Microsoft Corporation.
This posting is provided "AS IS" with no warranties, and confers no rights.
 
Karthi
I am experiencing pretty much the same symptoms as Naresh. I have 3 W2K servers with approx 110 people connecting to them using ICA clients on Winterms. I have the NLB set to Multicast Enabled. I have even reset their IP addresses to be released after 1 minute (where it was set for a 7 day lease). The affinity is set to Single. I find that the IP address reconnects to the last server it was connected to so it doens't go to the 'least used' server - which is what I would prefer. Should I set the affinity to None and what implications will that have
regard
Roger
 
Roger,
When affinity is set to Single, NLB uses the client IP address only to
determine which server will accept the connection. So, it is expected that
connections from a given client IP address will always go to the same
server. When you set the affinity to None, NLB will also use the client
TCP/UDP port (in addition to the client IP address) to decide on which
server should take the connection. As multiple ongoing connections from a
given client IP address would NOT have the same TCP/UDP port, the chances of
them taken by different servers are maximized. But, as I said, you are only
increasing the chances....NLB does not follow a deterministic algorithm like
least used or round-robin. However, given a large client population, you
should see even load balance. Please be aware that if the application hosted
on the NLB cluster relies on client session affinity where it assumes all
TCP connections from a given client IP address belonging to an ongoing
"application session" will go to the same server, setting Affinity to None
will break that. Let me know if this explanation is not clear and I will try
to present it differently. Feel free to ask questions.
--
Thanks,
Karthic
Software Design Engineer
Windows Server Group
Microsoft Corporation.
This posting is provided "AS IS" with no warranties, and confers no rights.

Roger said:
Karthic
I am experiencing pretty much the same symptoms as Naresh. I have 3 W2K
Click to expand...
servers with approx 110 people connecting to them using ICA clients on
Winterms. I have the NLB set to Multicast Enabled. I have even reset their
IP addresses to be released after 1 minute (where it was set for a 7 day
lease). The affinity is set to Single. I find that the IP address
reconnects to the last server it was connected to so it doens't go to the
'least used' server - which is what I would prefer. Should I set the
affinity to None and what implications will that have?
 
Hi Karthi

Thanks for getting back to me. I do understand most of what you've written. Not sure if you are recommending I do change the Affinity to none. Are you saying that some applications on the Terminal Servers may be affected by the Affinity setting? How do I determine this/which ones? If I change the Affinity to none, what happens if the session gets disconnected for whatever reason, would that Client restablish the connection to the same Server? Let me give you an example of what has happened only today... One of the three Servers reset itself unexpectedly (this is a different issue!) and people immediately started logging in to the other two running servers, to a point where I have 45 on one server and 46 on the other! The server that reset itself comes back-up and joins the cluster. I go to a Winterm that has made a connection to the cluster yesterday, not yet today, and when I connect to the cluster it goes to one of the two heavily used servers with 40+ users on it and not the one with only 1 or 2 people connected! This is the problem I am trying to get away from - where does the IP address or whatever of the Client get stored in the NLB and how can I flush this information so that a truly random and even spread of clients can take place. Ideally, I would prefer that anyone logging into the cluster after that server came back up, would go straight to it and not continually increase the load on the other two servers. Even if I ask people to log off from the heavily used servers and login to the cluster again, they will go straight back to the last server they were on, so there is no point in me going through that exercise and in the mean time more and more people are connecting to the heavily used servers, this in turn causes slowness and eventually brings the server down

Sorry about this long winded reply, but I want to give you as much info as I can, as I would really like to get a solution for this problem. Hope you can help

Regard
Roger
 
Back
Top