If this happens always, it literally means that the machine exists but that it has no services listening on the specified port, or there is a firewall stopping you. If it happens occasionally - you used the word "sometimes" - and retrying succeeds, it is likely because the server has a full 'backlog'. When you are waiting to be accepted on a listening socket, you are placed in a backlog. This backlog is finite and quite short - values of 1, 2 or 3 are not unusual - and so the OS might be unable to queue your request for the 'accept' to consume. The backlog is a parameter on the listen function - all languages and platforms have basically the same API in this regard, even the C# one. This parameter is often configurable if you control the server, and is likely read from some settings file or the registry. Investigate how to configure your server. If you wrote the server, you might have heavy processing in the accept of your socket, and this can be better moved to a separate worker-thread so your accept is always ready to receive connections. There are various architecture choices you can explore that mitigate queuing up clients and processing them sequentially. Regardless of whether you can increase the server backlog, you do need retry logic in your client code to cope with this issue - as even with a long backlog the server might be receiving lots of other requests on that port at that time. There is a rare possibility where a NAT router would give this error should it's ports for mappings be exhausted. I think we can discard this possibility as too much of a long shot though, since the router has 64K simultaneous connections to the same destination address/port before exhaustion.
有了backlog于是继续google关键字“WCF backlog”发现wcf binding配置确实有一个listenBacklog的项目，默认值是10,于是把服务的listenBacklog改成100，问题搞定。
对了添加listenBacklog属性的时候有个注意的是一定要移除一个默认的endpoint <endpoint address="mex" binding="mexTcpBinding" bindingConfiguration="" contract="IMetadataExchange" />这个endpoint是用来给vs等发现元数据用的，如果这个不移走启动服务的时候会报端口已经被监听的错误。