Pro, is your hardware load balancer really balanced?

Currently using a hardware load balancer as an Exchange deployment scenario, using a hardware load balancer a good advantage is that the load of the application can be distributed more evenly to the servers in each backend, there are two different modes of operation of the hardware load balancer, one is forwarding and the other is the proxy mode.

Forwarding using the mode of operation is very similar to Microsoft's NLB, the use of a stateless way to forward, when the service request came, I through the network detection to determine that the server does not have an outage, the message will be forwarded directly to the corresponding machine, and do not care about the corresponding process on this machine to provide services, There is no doubt that this model has some drawbacks. At the same time their gateways use the gateways of the current network and do not need to set up the gateway to a true load balancer device.

Based on the disadvantage of forwarding is because he works in the network layer, do not care about the specific application data corresponding to the server can receive, based on the existence of such a problem, there is a 7-tier load balancing device, that is, whether the application layer data can be properly accepted and processed, 7-tier load balancing device more work in the proxy mode. So based on this model we can detect the health status of the server. At the same time, according to the Port and service status to determine whether the server conforms to the provision of the appropriate service standards to decide whether to forward the service to the appropriate server. So his working mode is basically the following diagram:

In contrast, the agent model based on the application detection is well suited for the appropriate deployment in the enterprise, but if the corresponding algorithm is incorrect or the relevant settings are not OK, we find that the reputation of the server metrics is realized when the load is balanced. As shown in the diagram below, the access pattern we used was cylic, based on a pattern where our access is a load balancer that forwards the corresponding accesses directly to each server in a certain order. When we monitor the counter indicator, we find that the user connection is unbalanced, we mainly monitor three metrics of the Exchange server and are the metrics based on RPC CLIENT access:

1. Active user count Active users connection

2. Connection Count Connection Count

3. User Count users statistics

When we set up the cylic we found that the number of connections and the corresponding number of active users in a certain situation did not achieve a good load balance, so we need to adjust the algorithm. In the application load balancer, there is an algorithm to die based on the hash value of the corresponding adjustment algorithm, we use the hash at the same time after the disconnection of the cache adjustment is not retained after we can find that we can achieve a good real load balance.

So in the general case we deal with setting up the load balancer may also be to some extent to ensure that the application load Balancer device can work well, we need to continue to connect to the client connection status, and finally this picture is our application load balancer adjusted results, You can adjust the corresponding parameters according to the result. Here I put forward a number of things and ideas to pay attention to, you can follow these paths to really find the solution you need to solve the problem, I hope that the wrong can be helpful: