Splunk Completes Acquisition of Plumbr Learn more

To blog |

How to use Asynchronous Servlets to improve performance

October 10, 2013 by Nikita Salnikov-Tarnovski Filed under: Java Locked Threads

Since we have post this article we have received a lot of feedback about it. Based on this feedback we have updated the examples to make them easier to understand and, hopefully, clarify any doubts in their correctness.

This post is going to describe a performance optimization technique applicable to a common problem related to modern webapps. Applications nowadays are no longer just passively waiting for browsers to initiate requests, they want to start  communication themselves. A typical example would involve chat applications, auction houses etc – the common denominator being the fact that most of the time the connections with the browser are idle and wait for a certain event being triggered.

This type of applications has developed a problem class of their own, especially when facing heavy load. The symptoms include starving threads, suffering user interaction, staleness issues etc.

Based on recent experience with this type of apps under load, I thought it would be a good time to demonstrate a simple solution. After Servlet API 3.0 implementations became mainstream, the solution has become truly simple, standardized and elegant.

But before we jump into demonstrating the solution, we should understand the problem in greater detail. What could be easier for our readers than to explain the problem with the help of some source code:

@WebServlet(urlPatterns = "/BlockingServlet")
public class BlockingServlet extends HttpServlet {

  protected void doGet(HttpServletRequest request, HttpServletResponse response) {
    writeResponse(response, "OK");

  public static void waitForData() {
    try {
    } catch (InterruptedException e) {

The servlet above is an example of how an application described above could look like:

  • Every 2 seconds some event happens. E.g. stock quote arrives, chat is updated and so on.
  • End-user request arrives, announcing interest to monitor certain events
  • The thread is blocked until the next event arrives
  • Upon receiving the event, the response is compiled and sent back to the client

Let me explain this waiting aspect. We have some external event that happens every 2 seconds. When new request from end-user arrives, it has to wait some time between 0 and 2000ms until next event. In order to keep it simple, we have emulated this waiting part with a call to Thread.sleep() for random number of ms between 0 and 2000. So every request waits on average for 1 second.

Now – you might think this is a perfectly normal servlet. In many cases, you would be completely correct – there is nothing wrong with the code until the application faces significant load.

In order to simulate this load I created a fairly simple test with some help from JMeter, where I launched 2,000 threads, each running through 10 iterations of bombarding the application with requests to the /BlockedServlet. Running the test with the deployed servlet on an out-of-the-box Tomcat 8.0.30 I got the following results:

  • Average response time: 9,492 ms
  • Minimum response time: 205 ms
  • Maximum response time: 11,368 ms
  • Throughput: 195 requests/second

The default Tomcat configuration has got 200 worker threads which, coupled with the fact that the simulated work is replaced with the sleep cycle of average duration of 1000ms, explains nicely the throughput number – in each second the 200 threads should be able to complete 200 sleep cycles, 1 second on average each. Adding context switch costs on top of this, the achieved throughput of 195 requests/second is pretty close to our expectations.

The throughput itself would not look too bad for 99.9% of the applications out there. However, looking at the maximum and especially average response times the problem starts to look more serious. Getting the worst case response in 11 seconds instead of the expected 2 seconds is a sure way to annoy your users.

Let us now take a look at an alternative implementation, taking advantage of the Servlet API 3.0 asynchronous support:

@WebServlet(asyncSupported = true, value = "/AsyncServlet")
public class AsyncServlet extends HttpServlet {

  protected void doGet(HttpServletRequest request, HttpServletResponse response) throws ServletException, IOException {

  private static ScheduledExecutorService executorService = Executors.newScheduledThreadPool(1);

  static {
    executorService.scheduleAtFixedRate(AsyncServlet::newEvent, 0, 2, TimeUnit.SECONDS);

  private static void newEvent() {
    ArrayList clients = new ArrayList<>(queue.size());
    clients.parallelStream().forEach((AsyncContext ac) -> {
      ServletUtil.writeResponse(ac.getResponse(), "OK");

  private static final BlockingQueue queue = new ArrayBlockingQueue<>(20000);

  public static void addToWaitingList(AsyncContext c) {

This bit of code is a little more complex, so maybe before we start digging into solution details, I can outline that this solution performed ~5x better latency- and ~5x better throughput-wise. Equipped with the knowledge of such results, you should be more motivated to understand what is actually going on in the second example.

The servlet’s doGet itself looks truly simple. Two facts are worth outlining though, the first of which declares the servlet to support asynchronous method invocations:

@WebServlet(asyncSupported = true, value = "/AsyncServlet")

The second important aspect is hidden in the following method

  public static void addToWaitingList(AsyncContext c) {

in which the whole request processing consists only of storing an instance AsyncContext into some queue. This AsyncContext holds the request and response that were provided by the container and we can use later to answer the user request. Thus incoming requests are waiting for the notification – this could be an updated bid on the monitored auction or the next message in the group chat. The important aspect here is that servlet container’s threads have finished with doGet for the time being and are free to accept another requests.

Now, the notification arrives every 2 seconds and we have implemented this as a scheduled event, that calls the newEvent method every 2 seconds. When it arrives all the waiting tasks in the queue are processed by a single worker thread responsible for compiling and sending the responses. Instead of blocking hundreds of threads to wait behind the external notification, we achieved this in a lot simpler and cleaner way – batching the interest groups together and processing the requests in a single thread.

And the results speak for themselves – the very same test on the very same Tomcat 8.0.30 with default configuration resulted in the following:

  • Average response time: 1,875 ms
  • Minimum response time: 356 ms
  • Maximum response time: 2,326 ms
  • Throughput: 939 requests/second

The specific case here is small and synthetic, but similar improvements are achievable in the real-world applications.

Now, before you run to rewrite all your servlets to the asynchronous servlets – hold your horses for a minute. The solution works perfectly on a subset of usage cases, such as group chat notifications and auction house price alerts. You will most likely not benefit in the cases where the requests are waiting behind unique database queries being completed. So, as always, I must reiterate my favorite performance-related recommendation – measure everything. Do not guess anything.

Not sure if threads behave or are causing problems? Let Plumbr monitor your Java app and tell you if you need to change your code.

But on the occasions when the problem does fit the solution shape, I can only praise it. Besides the now-obvious improvements on throughput and latency, we have elegantly avoided possible thread starvation issues under heavy load.

Another important aspect – the approach to asynchronous request processing is finally standardized. Independent of your favorite Servlet API 3.0 – compliant application server such as Tomcat 7+, JBoss 6 or Jetty 8+ – you can be sure the approach works. No more wrestling with different Comet implementations or platform-dependent solutions, such as the Weblogic FutureResponseServlet.

Now, if you came with me this far, I can only recommend you to follow us on Twitter. I will keep posting on relevant and interesting tech topics, I promise.





Help with Jetty.
Hi hello, I have a concern,
I was testing Jetty and AsynContex, to simulate load uses Thread.sleep, and between test and test, note that the server does not serve more than 6 connections with AsynContex, if not until one of the connections responds, I would have idea of what’s that, which for me I think is a serious problem, I use Linux, Ubuntu Server 16, and java 8, and there’s nothing about it on the internet, please if you can help me !!

david colemanres

Ayuda con Jetty.
hey hola, tengo una preocupación,
Estaba probando Jetty y AsynContex, para simular carga use el Thread.sleep ,y entre prueba y prueba, note que el servidor, no atiende mas de 6 conexiones con AsynContex, si no hasta que una de las conexiones responde, tendran alguna idea de de por que pasa eso, lo cual para mi creo que es un problema grave, uso linux, Ubuntu Server 16, y java 8, y no encuentro nada al respecto en internet, por favor si puees ayudarme!!

david colemanres

It doesn’t seem sensible to have a 200-thread configuration if you’re bombarded by 2000 threads. I’m interested to know how it compares once you expand the pool size.


Do you suggest to have a thread per user?

Nikita Salnikov-Tarnovski

Perhaps. Or maybe something in between. Or an expandable thread pool.

I guess the real problem is that your example is too theoretical. If threads are doing nothing but waiting, you can basically create as many as you need. Only if threads need CPU time do you have to strike a balance to prevent thread starvation.


Let me disagree with you. You cannot use a thread-per-request model in any but the trivial applications. It is the common case to have much smaller number threads than actual users.


looking at your example, thank you btw. but what is ServletUtil.writeResponse(…) doing? I don’t see source for this class in example.


It just writes a response back to the client. The exact method is not important for the article’s main point.


Really confusing example. And second implementation logic is different. Let me explain. You have working thread that constantly running and checking queue every 2 sec and processing requests in the queue. Now, at the beginning your queue is empty and working thread falls asleep for 2 sec. Some requests arrives. After sleep, working thread starts processing the queue. But meanwhile new requests still arriving and getting to the queue. It MAY happen that queue is never empty again, and working thread will never call Thread.sleep(2000) again. Or will call it much less frequently. It may happen or may not. But anyway this 2 second notification mock implementation is wrong and may lead to wrong results.
Although I understand your intent and async processing will give performance gain, but this example implementation is bad.

Sec Ad

Hm. I think that you have a valid comment here. Thank you for bringing my attention to this! I will try to fix my example in the near future.


I have updated our async example. Hope it is more correct now. Thank you for your comment once more 🙂

Nikita Salnikov-Tarnovski Post author

Thank you for the post. Who will call contextInitialized() method?

Dinoop p

The servlet container will call it, as per Servlet API specification.


I’m sorry to disappoint you but this example is wrong :p In the first example sleep() executes every call while async implementation calls sleep once before processing queue in a loop.

By the way, your blog is great and has already learnt me a lot. Thank you for your effort.

Dzafar Sadik

Thank you for your kind words 🙂 But I still believe that second implementation is absolutely correct. We try to simulate the situation, where some events arrive every 2 seconds, e.g. brokerage info. When it arrives, every waiting request is notified at once, without further delays.

I have modified the blog post to make this clearer. 2 seconds are not the request processing time. It is time between events that requests are waiting for. When event arrives, further processing is very fast.

Nikita Salnikov-Tarnovski Post author

Why do you need the async context? Can’t it be done passing just the servlet response?


I beg your pardon, passing where?


to the Work class…implement the Work class like:
static void add(HttpServletResponse response){

Isn’t this the same? I don’t understand why there is the needs of the startAsync context


You cannot write into Response after servlet container has done with the current request processing. That is the whole point, why these async servlets were invented.


Ok…I read a little bit around and did some test…for what I read the async context is done for comet like stuff…I was looking if it can be useful for my stuff but no 🙂 thanks


Hello !

Good post, however your two implementations are not really the same.

In your blocking servlet, “backend process mock” (your Thread.sleep) takes 2 seconds.

In your async implementation, what takes 2 seconds is only a waiting time before two polling of the queue (and not the “backend process mock”

If you really wanted to have the same test, you should have your Thread.sleep into the while ((context = queue.poll()) != null) and not outside.

Can you take a look ?
I think your performance will sharply decrease, because one worker stacking process of 2 seconds won’t be fast enough.

while ((context = queue.poll()) != null) {
while ((context = queue.poll()) != null) {
while ((context = queue.poll()) != null) {


I still believe that second implementation is absolutely correct. We try to simulate the situation, where some events arrive every 2 seconds, e.g. brokerage info. When it arrives, every waiting request is notified at once, without further delays.

Buy may be our first implementation is not that correct… Right now every request waits for exactly 2 seconds, which is not as it should be… Thanks for pointing this out, I will try to find some time to clarify this.


Thanks for your quick reply.


Hi, i don’t know how to understand this sentence:

in each second the 200 threads should be able to complete 100 sleep cycles, 2 seconds each


Great. Just to test my understanding, isn’t this even easier to do with a WriteListener in Servlet 3.1? Or is that something different?

Bit Crusher

May be. Had no chance so far to learn that part of the API 🙂


great. I almost ignored Servlet 3.0 API. Will help in impressing others with my skills in next project (current one is like 5 year old app being supported with bug fixes and enhancements)

Java Experience

Same here. There is no much information floating around about what that Servlets 3 are all about and why should we care. We hoped to shed a little bit more light on it 🙂