In a client-server architecture, the client is always the one that initiates a connection, regardless of the type of hardware it is on. That means that in a case where a server of some kind 'initiates' a connection, for that connection it is the client and is expecting a response from its partner, the server.
This article explains why it is required to set acceleration policies on the client.
In an accelerated scenario, where packets are accelerated between two Exindas, the policy to accelerate traffic is required to be on the client side. If the policies to accelerate traffic were on the server side, the client Exinda - that is the Exinda closest to the initial request point - would have no knowledge of this, as policies are not shared amongst acceleration community members. That means that traffic would be all the way to the Exinda on the server side before attempting to do acceleration; at that point, there would be no Exinda to remove the acceleration and can result in some dropped connections at the server side.
If there are policies on the client-side Exinda to accelerate traffic, as long as the Exinda on the server side has an accelerated license installed, it does not need any corresponding policies. It will automatically recognize the traffic as accelerated and will do the necessary work to strip out the TCP options to prepare for the destination as long as the processes are running. In the same way, if the server responds in a connection that is accelerated, the server side Exinda will recognize it and accelerate the traffic.
Only if the server initiates a new connection will a policy explicitly need to be on its Exinda.