The number of simultaneous synchronous requests a server is capable of supporting is determined by the number of threads in the server’s thread pool (see
Section 32.10). If all of the threads are busy dispatching long-running operations, then no threads are available to process new requests and therefore clients may experience an unacceptable lack of responsiveness.
Asynchronous Method Dispatch (
AMD), the server-side equivalent of AMI (see
Section 10.17), addresses this scalability issue. Using AMD, a server can receive a request but then suspend its processing in order to release the dispatch thread as soon as possible. When processing resumes and the results are available, the server sends a response explicitly using a callback object provided by the Ice run time.
AMD is transparent to the client, that is, there is no way for a client to distinguish a request that, in the server, is processed synchronously from a request that is processed asynchronously.
In practical terms, an AMD operation typically queues the request data (i.e., the callback object and operation arguments) for later processing by an application thread (or thread pool). In this way, the server minimizes the use of dispatch threads and becomes capable of efficiently supporting thousands of simultaneous clients.
An alternate use case for AMD is an operation that requires further processing after completing the client’s request. In order to minimize the client’s delay, the operation returns the results while still in the dispatch thread, and then continues using the dispatch thread for additional work.
To enable asynchronous dispatch, you must add an ["amd"] metadata directive to your Slice definitions. The directive applies at the interface and the operation level. If you specify
["amd"] at the interface level, all operations in that interface use asynchronous dispatch; if you specify
["amd"] for an individual operation, only that operation uses asynchronous dispatch. In either case, the metadata directive
replaces synchronous dispatch, that is, a particular operation implementation must use synchronous or asynchronous dispatch and cannot use both.
["amd"] interface I {
bool isValid();
float computeRate();
};
interface J {
["amd"] void startProcess();
int endProcess();
};
In this example, both operations of interface I use asynchronous dispatch, whereas, for interface
J,
startProcess uses asynchronous dispatch and
endProcess uses synchronous dispatch.
Specifying metadata at the operation level (rather than at the interface or class level) minimizes the amount of generated code and, more importantly, minimizes complexity: although the asynchronous model is more flexible, it is also more complicated to use. It is therefore in your best interest to limit the use of the asynchronous model to those operations that need it, while using the simpler synchronous model for the rest.
1. A callback interface used by the implementation to notify the Ice run time about the completion of an operation. The name of this interface is formed using the pattern
AMD_class_op. For example, an operation named
foo defined in interface
I results in an interface named
AMD_I_foo. The interface is generated in the same scope as the interface or class containing the operation. Two methods are provided:
The ice_response method allows the server to report the successful completion of the operation. If the operation has a non-
void return type, the first parameter to
ice_response is the return value. Parameters corresponding to the operation’s
out parameters follow the return value, in the order of declaration.
The ice_exception method allows the server to raise an exception. With respect to exceptions, there is less compile-time type safety in an AMD implementation because there is no
throws clause on the dispatch method and any exception type could conceivably be passed to
ice_exception. However, the Ice run time validates the exception value using the same semantics as for synchronous dispatch (see
Section 4.10.4).
Neither ice_response nor
ice_exception throw any exceptions to the caller.
2.
The dispatch method, whose name has the suffix _async. This method has a
void return type. The first parameter is a reference to an instance of the callback interface described above. The remaining parameters comprise the
in parameters of the operation, in the order of declaration.
interface I {
["amd"] int foo(short s, out long l);
};
public interface AMD_I_foo {
void ice_response(int __ret, long l);
void ice_exception(java.lang.Exception ex);
}
void foo_async(AMD_I_foo __cb, short s);
There are two processing contexts in which the logical implementation of an AMD operation may need to report an exception: the dispatch thread (the thread that receives the invocation), and the response thread (the thread that sends the response).
1 Although we recommend that the callback object be used to report all exceptions to the client, it is legal for the implementation to raise an exception instead, but only from the dispatch thread.
As you would expect, an exception raised from a response thread cannot be caught by the Ice run time; the application’s run time environment determines how such an exception is handled. Therefore, a response thread must ensure that it traps all exceptions and sends the appropriate response using the callback object. Otherwise, if a response thread is terminated by an uncaught exception, the request may never be completed and the client might wait indefinitely for a response.
Whether raised in a dispatch thread or reported via the callback object, user exceptions are validated as described in
Section 4.10.2, and local exceptions may undergo the translation described in
Section 4.10.4.
module Demo {
sequence<float> Row;
sequence<Row> Grid;
exception RangeError {};
interface Model {
["amd"] Grid interpolate(Grid data, float factor)
throws RangeError;
};
};
Given a two-dimensional grid of floating point values and a factor, the interpolate operation returns a new grid of the same size with the values interpolated in some interesting (but unspecified) way.
Our servant class derives from Demo._ModelDisp and supplies a definition for the
interpolate_async method that creates a
Job to hold the callback object and arguments, and adds the
Job to a queue. The method is synchronized to guard access to the queue:
public final class ModelI extends Demo._ModelDisp {
synchronized public void interpolate_async(
Demo.AMD_Model_interpolate cb,
float[][] data,
float factor,
Ice.Current current)
throws RangeError
{
_jobs.add(new Job(cb, data, factor));
}
java.util.LinkedList _jobs = new java.util.LinkedList();
}
After queuing the information, the operation returns control to the Ice run time, making the dispatch thread available to process another request. An application thread removes the next
Job from the queue and invokes
execute, which uses
interpolateGrid (not shown) to perform the computational work:
class Job {
Job(Demo.AMD_Model_interpolate cb,
float[][] grid,
float factor)
{
_cb = cb;
_grid = grid;
_factor = factor;
}
void execute()
{
if (!interpolateGrid()) {
_cb.ice_exception(new Demo.RangeError());
return;
}
_cb.ice_response(_grid);
}
private boolean interpolateGrid() {
// ...
}
private Demo.AMD_Model_interpolate _cb;
private float[][] _grid;
private float _factor;
}
If interpolateGrid returns
false, then
ice_exception is invoked to indicate that a range error has occurred. The
return statement following the call to
ice_exception is necessary because
ice_exception does not throw an exception; it only marshals the exception argument and sends it to the client.
If interpolation was successful, ice_response is called to send the modified grid back to the client.
