EMA API Performance improvement

@dimple.shah

Thank you for reaching out to us.

You can check the Real-Time SDKs Performance Test Summary article that summarizes the performance report.

You can also refer to the EMA PerfTools in the package which demonstrate best practice and coding for performance.

The key point is that, in the EMA callback methods (such as onUpdateMsg), the application must return as fast as possible. The application can clone retrieved messages and then put them into queues so the messages can be processed by other threads. Therefore, a dispatching thread is freed to dispatch next incoming messages.

Otherwise, the application can use the horizontal scaling by creating multiple OMM Consumer instances. You can refer to the Horizontal Scaling example in the API package.

@dimple.shah

According to this discussion, there is no request throttle in EMA. This means that all requests registered by application will directly be sent to server. Moreover, the watchlist size used to generate the performance report is 100,000 items so 30k items should be okay in EMA.

You may run the EMA PerfTools in the package to verify the number of subcribed items. The PerfTools are uses to generate the performance report.

Are you using EMA C++, Java, or C#?

@dimple.shah

I modified the ex100_MP_Streaming example to subcribe to 30k items from the simulated service and found that it can retrieve all 30k refresh messages.

{

		OmmConsumer consumer = null;


		try

		{

			AppClient appClient = new AppClient();


			OmmConsumerConfig config = EmaFactory.createOmmConsumerConfig();


			consumer  = EmaFactory.createOmmConsumer(config.host("127.0.0.1:14002").username("user"));



			ReqMsg reqMsg = EmaFactory.createReqMsg();
for(int i=1;i<=30000;i++) {

			

				consumer.registerClient(reqMsg.serviceName("DIRECT_FEED").name("PTT"+i+".BK"), appClient);

			}



			while(true)

				Thread.sleep(60000);			

		} 

You may try modify the simple example to subscribe to 30k items and then check if it can retrieve all responses.

DIRECT_FEED is a simulated feed. I need to check that EMA Java can subscribe to 30k items.

You can run the EMA PerfTools for benchmarking.

@dimple.shah

To clone an update message in EMA Java, you can use the EmaFactory.createUpdateMsg​(UpdateMsg other) method. The cloned messages are put into a queue. Then, another thread gets a clone message from the queue and processes it.

Sorry, as far as I know, we don't have an example for it.

@dimple.shah

To improve the performance of EMA applications, there are two key points.

1. The EMA callback methods must return as fast as possible
2. The parameters passed to the callback methods can't be used outside the callback methods

If the application requires only some fields such as, BID and ASK, the application can use view requests to request only the BID and ASK field. The view requests can reduce the size of refresh and update messages and reduce the processing time.

Instead of using the EmaFactory to clone messages, the application can decode messages to extract required data (BID value and ASK value) and put these values to a queue to be processed outside the callback methods.

Regarding the performance of the EmaFactory.createRefreshMsg method, I have tested the code with the EMA example. On my machine, it takes milliseconds to clone a message.

Time taken to clone: 1ms

It should not take that long (8 seconds) to clone a message. Please contact the support team to investigate this issue.

If you have a developer connect contact (RDC), you can submit this issue to the API support team via Contact Premium Support. Otherwise, you can submit this issue on GitHub.