Asynchronous Client HTTP Transport
By default, CXF uses a transport based on the in-JDK HttpURLConnection object to perform HTTP requests. The HttpURLConnection object uses a blocking model for all IO operations which requires a per-thread execution model. From a pure performance standpoint, this model generally performs very well, but it does have problems scaling when many requests need to be executed simultaneously.
Also, the JAX-WS specification allows for generation of asynchronous methods on generated proxies as well as using asynchronous methods on the Dispatch objects. These methods can take an AsyncHandler object and return a polling Future object so applications do not have to wait for the response. With the HttpURLConnection based transport, CXF was forced to consume a background thread for each outstanding request.
CXF also has an HTTP client transport that is based on the Apache HTTP Components HttpAsyncClient library. Its Maven artifactId is cxf-rt-transports-http-hc. The HttpAsyncClient library uses a non-blocking IO model. This allows many more requests to be outstanding without consuming extra background threads. It also allows greater control over things like Keep-Alive handling which is very difficult or impossible with the HttpURLConnection based transport. However, the non-blocking model does not perform quite as well as the blocking model for pure synchronous request/response transactions.
By default, if the cxf-rt-transports-http-hc module is found on the classpath, CXF will use the HttpAsyncClient based implementation for any Async calls, but will continue to use the HttpURLConnection based transport for synchronous calls. This allows a good balance of performance for the common synchronous cases with scalability for the asynchronous cases. However, using a contextual property of "use.async.http.conduit" and set to true/false, you can control whether the async or blocking version is used. If "true", the HttpAsyncClient will be used even for synchronous calls, if "false", asynchronous calls will rely on the traditional method of using HTTPURLConnection along with a work queue to mimic the asynchronocity.
Another reason to use the asynchronous transport is to use HTTP methods that HttpURLConnection does not support. For example, the github.com REST API specifies the use of PATCH for some cases, but HttpURLConnection rejects PATCH.
Using the HTTP Components Transport from Java Code
To force global use of the HTTP Components transport, you can set a bus-level property:
The "normal" CXF/JAX-WS method of setting user credentials via the BindingProvider.USERNAME_PROPERTY/PASSWORD_PROPERTY will work with the Async transport as well. However, the HttpAsyncClient library does have some additional capabilities around NTLM that can be leveraged. In order to use that, you need to:
- Turn on the AutoRedirect and turn off the Chunking for the Conduit. This will allow CXF to cache the response in a manner that will allow the transport to keep resending the request during the authentication negotiation.
The Asynchronous HTTP Transport has several options that can set using Bus properties or via the OSGi configuration services to control various aspects of the underlying Apache HTTP Components HttpAsyncClient objects.
Settings related to the underlying TCP socket (see java.net.Socket for a definition of these values):
org.apache.cxf.transport.http.async.TCP_NODELAY (Default true)
Settings related to Keep-Alive connection management:
Maximum time a connection is held open in ms. Default is 60000.
Maximum number of connections opened per host. Default is 1000.
Maximum number of connections opened in total. Default is 5000.
Settings related to Apache HttpAsyncClient threads and selectors:
Number of threads HttpAsyncClient uses to process IO events. Default is "-1" which means one thread per CPU core.
true/false for whether the interest ops are queues or process directly.
Default 1000 ms. How often the selector thread wakes up if there are no events to process additional things like queue expirations.
Setting to control which conduit is used
ALWAYS, ASYNC_ONLY, NEVER.
Similar in meaning to the "use.async.http.conduit" context property described above. Whether to use the HttpAsyncClient: ALWAYS for both synchronous and asynchronous calls, ASYNC_ONLY (default) for asynchronous calls only, NEVER will use HTTPURLConnection for both types of calls.