In general when you are load balancing a cluster you can evenly spread the connections through the cluster and you get pretty consistent and even load balancing. However with some applications such as RDS (Microsoft Terminal Servers), you might get very high load from just a  few users doing heavy work. This in turn could result in slow performance for other users on the same server.

The solution to this problem is to use some kind of server load feedback agent.

The Loadbalancer.org appliance has had an adaptive feedback agent for a while now. However with a LOT of help from Simon Horman (of UltraMonkey fame) - we've managed to integrate the functionality into the main branch of HAproxy.

We also thought it would be a good idea to open source our previous work on Ldirectord/LVS, make it compatible with HAProxy, and release our Windows service code as GPL.

UPDATE: The Loadbalancer.org feedback agent code is now fully supported in HAProxy 1.5-dev21 + and all future versions.

How do I install the feedback agent on Windows?

Simply download the latest Windows Feedback Agent installation package here:
http://downloads.loadbalancer.org/agent/loadbalanceragent.msi (v4.5.3)

The feedback agent functionality is available in all versions of the Loadbalancer.org appliance software for both layer 4 (Ldirectord) and layer 7 (HAProxy).

Assuming that you have downloaded and installed the msi file from the link above, you should be able to find the simple client here:
C:\ProgramData\LoadBalancer.org\LoadBalancer\monitor.exe

Feedback

Simply hit the 'start' button and the agent should start responding to telnet on port 3333 (you may need to make an exception for that port in your Windows firewall).

Then make sure you modify your virtual server on the load balancer to poll the feedback agent:

FeedbackAgent

Then on the Windows server client:
You can change the 'mode' setting to drain then 'apply settings and restart' and HAProxy will then set the weight to 0 and status to drain (blue) i.e.:
drain

Or you can set the 'mode' to halt then 'apply settings and restart' and HAProxy will then immediately set the status to DOWN (yellow) i.e.:
down
When the agent is running in normal mode it will report back the percentage idle of the system based on the settings in the feedback agent XML file:

C:\ProgramData\LoadBalancer.org\LoadBalancer\config.xml

<xml>
  <Cpu>
    <ImportanceFactor value="1" />
    <ThresholdValue value="100" />
  </Cpu>
  <Ram>
    <ImportanceFactor value="0" />
    <ThresholdValue value="100" />
  </Ram>
  <TCPService>
    <Name value="HTTP" />
    <IPAddress value="*" />
    <Port value="80" />
    <MaxConnections value="0" />
    <ImportanceFactor value="0" />
  </TCPService>
  <ReadAgentStatusFromConfig value="False" />
  <ReadAgentStatusFromConfigInterval value="5" />
  <AgentStatus value="Normal" />
  <Interval value="10" />
</xml>

Notice that you can control both the importance of CPU & RAM utilization and also a threshold, so the following logic is used:

  • If CPU importance = 0 then ignore, 0.5 means give 50% importance, 1 means 100% importance
  • If RAM importance = 0 then ignore etc.

If ThresholdValue is reached on any monitor then immediately go into DRAIN mode (a value of 0 means no threshold is set).

This can be very useful if you have a small number of RDP sessions using a lot of RAM, simply set a ThresholdValue of 85, then as soon as memory crosses that threshold no new users will be sent to that server.

Otherwise to calculate the percentage idle reported by the agent would be to divide the utilization by the number of factors involved i.e.

If you are using two services then:

  • utilization = utilization + cpuLoad * cpuImportance%;
  • utilization = utilization + ramOccupied * ramImportance%;
  • utilization = utilization / 2

So if importance was 1 for both cpu and ram you would only get 0% reported if both CPU and RAM were 100%. (actually it gets a bit weird if one is already reporting 0%, but lets not worry about that....

And if the importance is zero then ignore completely i.e.

  • AS: utilization = utilization + cpuLoad * cpuImportance%;
  • IF: utilization = utilization + ramOccupied * 0 (importance is zero so ignore)
  • THEN: utilization = utilization (one service only so don't divide)

Also the final section TCPService effectively lets you load balance on number of established connections to your server, so you could balance based on the number of RDP connections to port 3389.

For this setting MaxConnections is important to specify as otherwise the agent will have no idea how to calculate the load i.e.

  • utilization = MaxConnections / 100 * number of current connections * importance%

In the following screen shot from a Loadbalancer.org appliance you can see that the Win2008R2 server is healthy and 99% idle, whereas the Linux server was busy at 43% idle before the Linux agent was put into maintenance mode and the server taken out of the group.

sysoverview

Does that make sense? Have a play with the config file and let us know what you think....

Ps. The Windows feedback agent source code is now available on GitHub

The agent functionality is obviously not just for Windows users:

Lets look at an example of load balancing Microsoft RDS with HAProxy:

listen RDSTest
bind 192.168.69.22:3389
mode tcp
balance leastconn
persist rdp-cookie
server backup 127.0.0.1:9081 backup non-stick
tcp-request inspect-delay 5s
tcp-request content accept if RDP_COOKIE
timeout client 12h
timeout server 12h
option tcpka
option redispatch
option abortonclose
maxconn 40000
server Win2008R2 192.168.64.50:3389 weight 100 check agent-check agent-port 3333 inter 2000 rise 2 fall 3 minconn 0 maxconn 0 on-marked-down shutdown-sessions

The important bit agent-check agent-port 3333 tells HAProxy to constantly monitor each backend server in the cluster by doing a telnet to port 3333 and grabbing the response which will usually be a percentage idle value i.e.

  • 80% - I am not very busy please increase my weight and send me more traffic
  • 10% - I'm busy please decrease my weight and stop sending me so much traffic
  • drain - Set the weight to 0 and gradually drain the traffic from this server for maintenance
  • stop - Stop all traffic immediately, kill this backend server
  • up ready 20% - Force HAProxy to bring the server up and set weight to 20% (irrespective of how it was taken down)

If you have a Linux backend you could create a simple service calling the following script:

#!/bin/bash
# This outputs a 1 second average CPU idle
LOAD=$(/usr/bin/vmstat 1 2 | /usr/bin/tail -1 | /usr/bin/awk '{ print $15 }' | /usr/bin/tee)
echo "$LOAD%"

Place the script into /usr/bin and call it

/usr/bin/lb-feedback.sh

Make sure that you make the script executable

chmod +x /usr/bin/lb-feedback.sh

Insert this line into /etc/services

lb-feedback 3333/tcp # loadbalancer.org feedback daemon

Now create the following file called /etc/xinetd.d/lb-feedback

# default: on
# description: lb-feedback socket server
service lb-feedback
{
port = 3333
socket_type = stream
flags = REUSE
wait = no
user = nobody
server = /usr/bin/lb-feedback.sh
log_on_success += USERID
log_on_failure += USERID
disable = no
}

Then change permissions and restart xinetd:

chmod 644 /etc/xinetd.d/lb-feedback
/etc/init.d/xinetd restart

You can now test this service by using telnet:

telnet 127.0.0.1 3333
Trying 127.0.0.1...
Connected to 127.0.0.1.
Escape character is '^]'.
95%
Connection closed by foreign host.