We Keep Coding

Data structure to manage a maximum number of requests per minute for an api

I need to send data to an external api, but this API has a limit of requests per endpoint(i.e: 60 requests per minute).
The data come from Kafka, then every message goes to redis(because I can send a request with 200 items). So, I use a simple cache to help me, and I can guarantee that if my server goes down, I wont lose any message.
The problem is, that there are moments when the Kafka starts to send to many messages, then the redis starts to grow(more than 1 million of messages to send to the api), and we can not make requests too fast as messages come in. Then, we have a big delay.
My first code was simple: ExecutorService executor = Executors.newFixedThreadPool(1);
This works very well, when there are few messages, and the delay is minimal.
So, the first thing that I did was change the executor to: ExecutorService executor = Executors.newCachedThreadPool();
So I can demand new threads as I need to make the requests to the external api faster, but, I have the problem with the limit of requests per minute.
There are endpoints that I can make 300 requests per minutes, others 500, others 30 and so on.
The code that I did is not very good, and this is for the company that I work, so, I really need to make this better.
So, every time that I am going to request the external api, I call the makeRequest method, this method is synchronized, I know that I could use a synchonized list, but I think that a synchronized method is better at this situation.
// This is an inner class
private static class IntegrationType {
final Queue<Long> requests; // This queue is used to store the timestamp of the requests
final int maxRequestsPerMinute; // How many requests I can make per minute
public IntegrationType(final int maxRequestsPerMinute) {
this.maxRequestsPerMinute = maxRequestsPerMinute;
this.requests = new LinkedList<>();
}
synchronized void makeRequest() {
final long current = System.currentTimeMillis();
requests.add(current);
if (requests.size() >= maxRequestsPerMinute) {
long first = requests.poll(); // gets the first request
// The difference between the current request and the first request of the queue
final int differenceInSeconds = (int) (current - first) / 1000;
// if the difference is less than the maximum allowed
if (differenceInSeconds <= 60) {
// seconds to sleep.
final int secondsToSleep = 60 - differenceInSeconds;
sleep(secondsToSleep);
}
}
}
void sleep( int seconds){
try {
Thread.sleep(seconds * 1000);
} catch (InterruptedException e) {
e.printStackTrace();
}
}
}
So, there is a Data Structure that I could use?
What considerations should I take?
Thanks in advance.

If understand your problem correctly, you can use a BlockingQueue with a ScheduledExecutorService as follows.
BlockingQueues have the method put which will only add the given element at the queue if there is available space, otherwise the method call will wait (until there is free space). They also have the method take which will only remove an element from the queue if there are any elements at all, otherwise the method call will wait (until there is at least one element to take).
Specifically you can use a LinkedBlockingQueue or an ArrayBlockingQueue which can be given with a fixed size of elements to hold at any given time. This fixed size means that you can submit with put as many requests as you like, but you will only take requests and process them once every second or something (so as to make 60 requests per minute for example).
To instantiate a LinkedBlockingQueue with fixed size, just use the corresponding constructor (which accepts the size as the argument). LinkedBlockingQueue will take elements in FIFO order according to its documentation.
To instantiate an ArrayBlockingQueue with fixed size, use the constructor which accepts the size but also the boolean flag named fair. If this flag is true then the queue will take elements also in FIFO order.
Then you can have a ScheduledExecutorService (instead of waiting inside a loop) where you can submit a single Runnable which will take from the queue, make the communication with the external API and then wait for the required delay between communications.
Follows a simple demonstration example of the above:
import java.util.Objects;
import java.util.concurrent.ArrayBlockingQueue;
import java.util.concurrent.BlockingQueue;
import java.util.concurrent.ExecutorService;
import java.util.concurrent.Executors;
import java.util.concurrent.TimeUnit;
public class Main {
public static class RequestSubmitter implements Runnable {
private final BlockingQueue<Request> q;
public RequestSubmitter(final BlockingQueue<Request> q) {
this.q = Objects.requireNonNull(q);
}
#Override
public void run() {
try {
q.put(new Request()); //Will block until available capacity.
}
catch (final InterruptedException ix) {
System.err.println("Interrupted!"); //Not expected to happen under normal use.
}
}
}
public static class Request {
public void make() {
try {
//Let's simulate the communication with the external API:
TimeUnit.MILLISECONDS.sleep((long) (Math.random() * 100));
}
catch (final InterruptedException ix) {
//Let's say here we failed to communicate with the external API...
}
}
}
public static class RequestImplementor implements Runnable {
private final BlockingQueue<Request> q;
public RequestImplementor(final BlockingQueue<Request> q) {
this.q = Objects.requireNonNull(q);
}
#Override
public void run() {
try {
q.take().make(); //Will block until there is at least one element to take.
System.out.println("Request made.");
}
catch (final InterruptedException ix) {
//Here the 'taking' from the 'q' is interrupted.
}
}
}
public static void main(final String[] args) throws InterruptedException {
/*The following initialization parameters specify that we
can communicate with the external API 60 times per 1 minute.*/
final int maxRequestsPerTime = 60;
final TimeUnit timeUnit = TimeUnit.MINUTES;
final long timeAmount = 1;
final BlockingQueue<Request> q = new ArrayBlockingQueue<>(maxRequestsPerTime, true);
//final BlockingQueue<Request> q = new LinkedBlockingQueue<>(maxRequestsPerTime);
//Submit some RequestSubmitters to the pool...
final ExecutorService pool = Executors.newFixedThreadPool(100);
for (int i = 0; i < 50_000; ++i)
pool.submit(new RequestSubmitter(q));
System.out.println("Serving...");
//Find out the period between communications with the external API:
final long delayMicroseconds = TimeUnit.MICROSECONDS.convert(timeAmount, timeUnit) / maxRequestsPerTime;
//We could do the same with NANOSECONDS for more accuracy, but that would be overkill I think.
//The most important line probably:
Executors.newSingleThreadScheduledExecutor().scheduleWithFixedDelay(new RequestImplementor(q), 0L, delayMicroseconds, TimeUnit.MICROSECONDS);
}
}
Note that I used scheduleWithFixedDelay and not scheduleAtFixedRate. You can see in their documentation that the first one will wait for the delay between the end of the call of the submitted Runnable to start the next one, while the second one will not wait and just resubmit the Runnable every period time units. But we don't know how long does it take to communicate with the external API, so what if for example we scheduleAtFixedRate with a period of once every minute, but the request takes more than a minute to be completed?... Then a new request would be submitted while the first one is not yet finished. So that is why I used scheduleWithFixedDelay instead of scheduleAtFixedRate. But there is more: I used a single thread scheduled executor service. Does that mean that if the first call is not finished, then a second cannot be started?... Well it seems, if you take a look at the implementation of Executors#newSingleThreadScheduledExecutor(), that a second call may occur because single thread core pool size, does not mean that the pool is of fixed size.
Another reason that I used scheduleWithFixedDelay is because of underflow of requests. For example what about the queue being empty? Then the scheduling should also wait and not submit the Runnable again.
On the other hand, if we use scheduleWithFixedDelay, with say a delay of 1/60f seconds between scheduling, and there are submitted more than 60 requests in a single minute, then this will surely make our throughput to the external API drop, because with scheduleWithFixedDelay we can guarantee that at most 60 requests will be made to the external API. It can be less than that, but we don't want it to be. We would like to reach the limit every single time. If that's not a concern to you, then you can use the above implementation already.
But let's say you do care to reach as close to the limit as possible every time, in which case and as far as I know, you can do this with a custom scheduler, which would be less clean solution than the first, but more time accurate.
Bottomline, with the above implementation, you need to make sure that the communication with the external API to serve the requests is as fast as possible.
Finaly, I should warn you to consider that I couldn't find what happens if the BlockingQueue implementations I suggested are not puting in FIFO order. I mean, what if 2 requests arrive at almost the same time while the queue is full? They will both wait, but will the first one which arrived be waiting and get puted first, or the second one be puted first? I don't know. If you don't care about some requests being made at the external API out of order, then don't worry and use the code up to this point. If you do care however, and you are able to put for example a serial number at each request, then you can use a PriorityQueue after the BlockingQueue, or even experiment with PriorityBlockingQueue (which is unfortunately unbounded). That would complicate things even more, so I didn't post relevant code with the PriorityQueue. At least I did my best and I hope I shed some good ideas. I am not saying this post is a complete solution to all your problems, but it is some considerations to start with.

I implemented something different that what #gthanop suggested.
Something that I discover, is that the limits may change. So, I might need to grow or shrink the blocking list. Another reason, would not be so easily to adapt our current code to this. And another one, we might use more than one instance, so we will need a distributed lock.
So, I implement something more easily, but not so efficiently as the answer of #ghtanop.
Here is my code(adapted, cause I can not show the company code):
import java.util.concurrent.Executors;
import java.util.concurrent.TimeUnit;
import java.util.concurrent.ScheduledExecutorService;
import java.util.HashMap;
import java.util.Map;
import java.util.concurrent.ConcurrentHashMap;
import java.util.concurrent.Semaphore;
public class Teste {
private static enum ExternalApi {
A, B, C;
}
private static class RequestManager {
private long firstRequest; // First request in one minute
// how many request have we made
private int requestsCount = 0;
// A timer thread, it will execute at every minute, it will refresh the request count and the first request time
private final ScheduledExecutorService executor = Executors.newScheduledThreadPool(1);
RequestManager() {
final long initialDelay = 0L;
final long fixedRate = 60;
executor.scheduleAtFixedRate(() -> {
System.out.println("Clearing the current count!");
requestsCount = 0;
firstRequest = System.currentTimeMillis();
}, initialDelay, fixedRate, TimeUnit.SECONDS);
}
void incrementRequest() {
requestsCount++;
}
long getFirstRequest() {
return firstRequest;
}
boolean requestsExceeded(final int requestLimit) {
return requestsCount >= requestLimit;
}
}
public static class RequestHelper {
private static final byte SECONDS_IN_MINUTE = 60;
private static final short MILLISECONDS_IN_SECOND = 1000;
private static final byte ZERO_SECONDS = 0;
// Table to support the time, and count of the requests
private final Map<Integer, RequestManager> requests;
// Table that contains the limits of each type of request
private final Map<Integer, Integer> requestLimits;
/**
* We need an array of semaphores, because, we might lock the requests for ONE, but not for TWO
*/
private final Semaphore[] semaphores;
private RequestHelper(){
// one semaphore for type
semaphores = new Semaphore[ExternalApi.values().length];
requests = new ConcurrentHashMap<>();
requestLimits = new HashMap<>();
for (final ExternalApi type : ExternalApi.values()) {
// Binary semaphore, must be fair, because we are updating things.
semaphores[type.ordinal()] = new Semaphore(1, true);
}
}
/**
* When my token expire, I must update this, because the limits might change.
* #param limits the new api limits
*/
protected void updateLimits(final Map<ExternalApi, Integer> limits) {
limits.forEach((key, value) -> requestLimits.put(key.ordinal(), value));
}
/**
* Increments the counter for the type of the request,
* Using the mutual exclusion lock, we can handle and block other threads that are trying to
* do a request to the api.
* If the incoming requests are going to exceed the maximum, we will make the thread sleep for N seconds ( 60 - time since first request)
* since we are using a Binary Semaphore, it will block incoming requests until the thread that is sleeping, wakeup and release the semaphore lock.
*
* #param type of the integration, Supp, List, PET etc ...
*/
protected final void addRequest(final ExternalApi type) {
// the index of this request
final int requestIndex = type.ordinal();
// we get the permit for the semaphore of the type
final Semaphore semaphore = semaphores[requestIndex];
// Try to acquire a permit, if no permit is available, it will block until one is available.
semaphore.acquireUninterruptibly();
///gets the requestManager for the type
final RequestManager requestManager = getRequest(requestIndex);
// increments the number of requests
requestManager.incrementRequest();
if (requestManager.requestsExceeded(requestLimits.get(type.ordinal()))) {
// the difference in seconds between a minute - the time that we needed to reach the maximum of requests
final int secondsToSleep = SECONDS_IN_MINUTE - (int) (System.currentTimeMillis() - requestManager.getFirstRequest()) / MILLISECONDS_IN_SECOND;
// We reached the maximum in less than a minute
if (secondsToSleep > ZERO_SECONDS) {
System.out.printf("We reached the maximum of: %d per minute by: %s. We must wait for: %d before make a new request!\n", requestLimits.get(type.ordinal()), type.name(), secondsToSleep);
sleep(secondsToSleep * MILLISECONDS_IN_SECOND);
}
}
// releases the semaphore
semaphore.release();
}
private final void sleep(final long time) {
try {
Thread.sleep(time);
} catch (InterruptedException e) {
e.printStackTrace();
}
}
/**
* Gets the first Request Manager, if it is the first request, it will create the
* RequestManager object
* #param index
* #return a RequestManager instance
*/
private RequestManager getRequest(final int index) {
RequestManager request = requests.get(index);
if(request == null) {
request = new RequestManager();
requests.put(index, request);
}
return request;
}
}
public static void main(String[] args) {
final RequestHelper requestHelper = new RequestHelper();
final Map<ExternalApi, Integer> apiLimits = Map.of(ExternalApi.A, 30, ExternalApi.B, 60, ExternalApi.C, 90);
// update the limits
requestHelper.updateLimits(apiLimits);
final ScheduledExecutorService executor = Executors.newScheduledThreadPool(3);
executor.scheduleWithFixedDelay(() -> {
System.out.println("A new request is going to happen");
requestHelper.addRequest(ExternalApi.A);
sleep(65);
}, 0, 100, TimeUnit.MILLISECONDS);
executor.scheduleWithFixedDelay(() -> {
System.out.println("B new request is going to happen");
requestHelper.addRequest(ExternalApi.B);
sleep(50);
}, 0, 200, TimeUnit.MILLISECONDS);
executor.scheduleWithFixedDelay(() -> {
System.out.println("C new request is going to happen");
requestHelper.addRequest(ExternalApi.C);
sleep(30);
}, 0, 300, TimeUnit.MILLISECONDS);
}
private static final void sleep(final long time) {
try {
Thread.sleep(time);
} catch (InterruptedException e) {
e.printStackTrace();
}
}
}

Lock free solution while updating the map and reading it

I am trying to measure how much time each thread takes in inserting to database. I have captured all those performance numbers in a ConcurrentHashMap named histogram like how much time each thread is taking in inserting.
Below is the code in which I am measuring how much time each thread is taking and storing it in a ConcurrentHashMap
class Task implements Runnable {
public static ConcurrentHashMap<Long, AtomicLong> histogram = new ConcurrentHashMap<Long, AtomicLong>();
#Override
public void run() {
try {
long start = System.nanoTime();
preparedStatement.executeUpdate(); // flush the records.
long end = System.nanoTime() - start;
final AtomicLong before = histogram.putIfAbsent(end / 1000000, new AtomicLong(1L));
if (before != null) {
before.incrementAndGet();
}
}
}
}
So my question is whether the way I am trying to measure how much time each thread is taking and storing all those numbers in a ConcurrentHashMap will be thread safe or not?
I think my whole update operation is Atomic. And I just wanted to see if there are any better approach than this if my whole operation is not Atomic. I am looking for mostly lock free solution.
And then after every thread is finished executing there tasks, I am printing this Histogram map from the main method as I have made that map as Static. So this way is right or not?
public class LoadTest {
public static void main(String[] args) {
//executing all the threads using ExecutorService
//And then I am printing out the historgram that got created in Task class
System.out.println(Task.histogram);
}
}

Your code is correct; there is also a (more complex) idiom that avoids instantiating AtomicLong each time. However do note that a "naïve" lock-based solution would probably be just as good due to a very low duration of the critical section.

Metrics from multiple threads

So this seems like a pretty common use case, and maybe I'm over thinking it, but I'm having an issue with keeping centralized metrics from multiple threads. Say I have multiple worker threads all processing records and I every 1000 records I want to spit out some metric. Now I could have each thread log individual metrics, but then to get throughput numbers, but I'd have to add them up manually (and of course time boundaries won't be exact). Here's a simple examples:
public class Worker implements Runnable {
private static int count = 0;
private static long processingTime = 0;
public void run() {
while (true) {
...get record
count++;
long start = System.currentTimeMillis();
...do work
long end = System.currentTimeMillis();
processingTime += (end-start);
if (count % 1000 == 0) {
... log some metrics
processingTime = 0;
count = 0;
}
}
}
}
Hope that makes some sense. Also I know the two static variables will probably be AtomicInteger and AtomicLong . . . but maybe not. Interested in what kinds of ideas people have. I had thought about using Atomic variables and using a ReeantrantReadWriteLock - but I really don't want the metrics to stop the processing flow (i.e. the metrics should have very very minimal impact on the processing). Thanks.

Offloading the actual processing to another thread can be a good idea. The idea is to encapsulate your data and hand it off to a processing thread quickly so you minimize impact on the threads that are doing meaningful work.
There is a small handoff contention, but that cost is usually a lot smaller than any other type of synchronization that it should be a good candidate in many situations. I think M. Jessup's solution is pretty close to mine, but hopefully the following code illustrates the point clearly.
public class Worker implements Runnable {
private static final Metrics metrics = new Metrics();
public void run() {
while (true) {
...get record
long start = System.currentTimeMillis();
...do work
long end = System.currentTimeMillis();
// process the metric asynchronously
metrics.addMetric(end - start);
}
}
private static final class Metrics {
// a single "background" thread that actually handles
// processing
private final ExecutorService metricThread =
Executors.newSingleThreadExecutor();
// data (no synchronization needed)
private int count = 0;
private long processingTime = 0;
public void addMetric(final long time) {
metricThread.execute(new Runnable() {
public void run() {
count++;
processingTime += time;
if (count % 1000 == 0) {
... log some metrics
processingTime = 0;
count = 0;
}
}
});
}
}
}

I would suggest if you don't want the logging to interfere with the processing, you should have a separate log worker thread and have your processing threads simply provide some type of value object that can be handed off. In the example I choose a LinkedBlockingQueue since it has the ability to block for an insignificant amount of time using offer() and you can defer the blocking to another thread that pulls the values from a queue. You might need to have increased logic in the MetricProcessor to order data, etc depending on your requirements, but even if it is a long running operation it wont keep the VM thread scheduler from restarting the real processing threads in the mean time.
public class Worker implements Runnable {
public void run() {
while (true) {
... do some stuff
if (count % 1000 == 0) {
... log some metrics
if(MetricProcessor.getInstance().addMetrics(
new Metrics(processingTime, count, ...)) {
processingTime = 0;
count = 0;
} else {
//the call would have blocked for a more significant
//amount of time, here the results
//could be abandoned or just held and attempted again
//as a larger data set later
}
}
}
}
}
public class WorkerMetrics {
...some interesting data
public WorkerMetrics(... data){
...
}
...getter setters etc
}
public class MetricProcessor implements Runnable {
LinkedBlockingQueue metrics = new LinkedBlockingQueue();
public boolean addMetrics(WorkerMetrics m) {
return metrics.offer(m); //This may block, but not for a significant amount of time.
}
public void run() {
while(true) {
WorkMetrics m = metrics.take(); //wait here for something to come in
//the above call does all the significant blocking without
//interrupting the real processing
...do some actual logging, aggregation, etc of the metrics
}
}
}

If you depend on the state of count and the state of processingTime to be in synch then you would have to be using a Lock. For example if when ++count % 1000 == 0 is true, you want to evaluate the metrics of processingTime at THAT time.
For that case, it would make sense to use a ReentrantLock. I wouldn't use a RRWL because there isn't really an instance where a pure read is occuring. It is always a read/write set. But you would need to Lock around all of
count++
processingTime += (end-start);
if (count % 1000 == 0) {
... log some metrics
processingTime = 0;
count = 0;
}
Whether or not count++ is going to be at that location, you will need to lock around that also.
Finally if you are using a Lock, you do not need an AtomicLong and AtomicInteger. It just adds to the overhead and isn't more thread-safe.

Counting the number of games played in 1 second using Java Threads

I want to know how many games my computer can play in 1000 ms. I did the tests before without using Threads (it plays 13k). Now that I think I'm using threads, I still get the same. Since I don't have much experience with Java threads, I assume I'm doing something wrong but I just can't get it.
Thanks in advance
public class SpeedTest<T extends BoardGame> implements Runnable
{
public static int gamesPlayed = 0;
private ElapsedTimer timer;
private double maxTime;
private BoardAgent<T> agent;
private BoardGame<T> game;
public SpeedTest(BoardGame<T> game, ElapsedTimer timer, double maxTime, Random rng)
{
this.game = game;
this.timer = timer;
this.maxTime = maxTime;
this.agent = new RandomAgent<T>(rng);
}
#Override
public void run()
{
while (true)
{
BoardGame<T> newBoard = game.copy();
while (!newBoard.isGameOver())
newBoard.makeMove(agent.move(newBoard));
gamesPlayed++;
if (timer.elapsedMilliseconds() > maxTime) {
break;
}
}
}
public static void main(String[] args)
{
Random rng = new Random();
BoardGame<Connect4> game = new Connect4(6, 7);
double maxTime = 1000;
ElapsedTimer timer = new ElapsedTimer();
SpeedTest<Connect4> speedTest1 = new SpeedTest<Connect4>(game, timer, maxTime, rng);
SpeedTest<Connect4> speedTest2 = new SpeedTest<Connect4>(game, timer, maxTime, rng);
Thread t1 = new Thread(speedTest1);
Thread t2 = new Thread(speedTest2);
t1.start();
t2.start();
try {
Thread.sleep((long) maxTime);
} catch (InterruptedException e) {
e.printStackTrace();
}
System.out.println("Games: " + SpeedTest.gamesPlayed);
}
}

I suspect that the reason that you are not seeing any speedup is that your application is only using 1 physical processor. If it is only using one processor, then the two threads won't be running in parallel. Instead, the processor will be "time-slicing" between the two threads.
What can you do about this?
Run on a dual-core etc processor. Or if you have a single processor machine with HT support, enable HT.
Run the test over a longer time; e.g. a number of minutes.
The reason I suggest the latter is that this could be a JVM warmup effect. When a JVM starts a new application, it needs to do a lot of class loading and JIT compilation behind the scenes. These tasks will be largely (if not totally) single-threaded. Running the tests over a longer period of time reduces the contribution of the "warm up" overheads to the average time per "game".
There is a fix that you ought to make to make the program thread-safe. Change
public static int gamesPlayed = 0;
to
private static final AtomicInteger gamesPlayed = new AtomicInteger();
and then use getAndIncrement() to increment the counter and intValue() to fetch its value. (This is simpler than having each thread maintain its own counter and summing them at the end.)
However, I strongly suspect that this change (or #Erik's alternative) will make little difference to the results you are seeing. I'm now sure it is either:
JVM warmup issue as described above,
a consequence of high object creation rates and/or heap starvation, or
some hidden synchronization issue between the instances of your game.

Don't use a static int, use a normal member int.
Instead of the sleep, call .join on both threads.
Then finally add the member ints.

ExecutorService's surprising performance break-even point --- rules of thumb?

I'm trying to figure out how to correctly use Java's Executors. I realize submitting tasks to an ExecutorService has its own overhead. However, I'm surprised to see it is as high as it is.
My program needs to process huge amount of data (stock market data) with as low latency as possible. Most of the calculations are fairly simple arithmetic operations.
I tried to test something very simple: "Math.random() * Math.random()"
The simplest test runs this computation in a simple loop. The second test does the same computation inside a anonymous Runnable (this is supposed to measure the cost of creating new objects). The third test passes the Runnable to an ExecutorService (this measures the cost of introducing executors).
I ran the tests on my dinky laptop (2 cpus, 1.5 gig ram):
(in milliseconds)
simpleCompuation:47
computationWithObjCreation:62
computationWithObjCreationAndExecutors:422
(about once out of four runs, the first two numbers end up being equal)
Notice that executors take far, far more time than executing on a single thread. The numbers were about the same for thread pool sizes between 1 and 8.
Question: Am I missing something obvious or are these results expected? These results tell me that any task I pass in to an executor must do some non-trivial computation. If I am processing millions of messages, and I need to perform very simple (and cheap) transformations on each message, I still may not be able to use executors...trying to spread computations across multiple CPUs might end up being costlier than just doing them in a single thread. The design decision becomes much more complex than I had originally thought. Any thoughts?
import java.util.concurrent.ExecutorService;
import java.util.concurrent.Executors;
import java.util.concurrent.TimeUnit;
public class ExecServicePerformance {
private static int count = 100000;
public static void main(String[] args) throws InterruptedException {
//warmup
simpleCompuation();
computationWithObjCreation();
computationWithObjCreationAndExecutors();
long start = System.currentTimeMillis();
simpleCompuation();
long stop = System.currentTimeMillis();
System.out.println("simpleCompuation:"+(stop-start));
start = System.currentTimeMillis();
computationWithObjCreation();
stop = System.currentTimeMillis();
System.out.println("computationWithObjCreation:"+(stop-start));
start = System.currentTimeMillis();
computationWithObjCreationAndExecutors();
stop = System.currentTimeMillis();
System.out.println("computationWithObjCreationAndExecutors:"+(stop-start));
}
private static void computationWithObjCreation() {
for(int i=0;i<count;i++){
new Runnable(){
#Override
public void run() {
double x = Math.random()*Math.random();
}
}.run();
}
}
private static void simpleCompuation() {
for(int i=0;i<count;i++){
double x = Math.random()*Math.random();
}
}
private static void computationWithObjCreationAndExecutors()
throws InterruptedException {
ExecutorService es = Executors.newFixedThreadPool(1);
for(int i=0;i<count;i++){
es.submit(new Runnable() {
#Override
public void run() {
double x = Math.random()*Math.random();
}
});
}
es.shutdown();
es.awaitTermination(10, TimeUnit.SECONDS);
}
}

Using executors is about utilizing CPUs and / or CPU cores, so if you create a thread pool that utilizes the amount of CPUs at best, you have to have as many threads as CPUs / cores.
You are right, creating new objects costs too much. So one way to reduce the expenses is to use batches. If you know the kind and amount of computations to do, you create batches. So think about thousand(s) computations done in one executed task. You create batches for each thread. As soon as the computation is done (java.util.concurrent.Future), you create the next batch. Even the creation of new batches can be done in parralel (4 CPUs -> 3 threads for computation, 1 thread for batch provisioning). In the end, you may end up with more throughput, but with higher memory demands (batches, provisioning).
Edit: I changed your example and I let it run on my little dual-core x200 laptop.
provisioned 2 batches to be executed
simpleCompuation:14
computationWithObjCreation:17
computationWithObjCreationAndExecutors:9
As you see in the source code, I took the batch provisioning and executor lifecycle out of the measurement, too. That's more fair compared to the other two methods.
See the results by yourself...
import java.util.List;
import java.util.Vector;
import java.util.concurrent.ExecutorService;
import java.util.concurrent.Executors;
import java.util.concurrent.TimeUnit;
public class ExecServicePerformance {
private static int count = 100000;
public static void main( String[] args ) throws InterruptedException {
final int cpus = Runtime.getRuntime().availableProcessors();
final ExecutorService es = Executors.newFixedThreadPool( cpus );
final Vector< Batch > batches = new Vector< Batch >( cpus );
final int batchComputations = count / cpus;
for ( int i = 0; i < cpus; i++ ) {
batches.add( new Batch( batchComputations ) );
}
System.out.println( "provisioned " + cpus + " batches to be executed" );
// warmup
simpleCompuation();
computationWithObjCreation();
computationWithObjCreationAndExecutors( es, batches );
long start = System.currentTimeMillis();
simpleCompuation();
long stop = System.currentTimeMillis();
System.out.println( "simpleCompuation:" + ( stop - start ) );
start = System.currentTimeMillis();
computationWithObjCreation();
stop = System.currentTimeMillis();
System.out.println( "computationWithObjCreation:" + ( stop - start ) );
// Executor
start = System.currentTimeMillis();
computationWithObjCreationAndExecutors( es, batches );
es.shutdown();
es.awaitTermination( 10, TimeUnit.SECONDS );
// Note: Executor#shutdown() and Executor#awaitTermination() requires
// some extra time. But the result should still be clear.
stop = System.currentTimeMillis();
System.out.println( "computationWithObjCreationAndExecutors:"
+ ( stop - start ) );
}
private static void computationWithObjCreation() {
for ( int i = 0; i < count; i++ ) {
new Runnable() {
#Override
public void run() {
double x = Math.random() * Math.random();
}
}.run();
}
}
private static void simpleCompuation() {
for ( int i = 0; i < count; i++ ) {
double x = Math.random() * Math.random();
}
}
private static void computationWithObjCreationAndExecutors(
ExecutorService es, List< Batch > batches )
throws InterruptedException {
for ( Batch batch : batches ) {
es.submit( batch );
}
}
private static class Batch implements Runnable {
private final int computations;
public Batch( final int computations ) {
this.computations = computations;
}
#Override
public void run() {
int countdown = computations;
while ( countdown-- > -1 ) {
double x = Math.random() * Math.random();
}
}
}
}

This is not a fair test for the thread pool for following reasons,
You are not taking advantage of the pooling at all because you only have 1 thread.
The job is too simple that the pooling overhead can't be justified. A multiplication on a CPU with FPP only takes a few cycles.
Considering following extra steps the thread pool has to do besides object creation and the running the job,
Put the job in the queue
Remove the job from queue
Get the thread from the pool and execute the job
Return the thread to the pool
When you have a real job and multiple threads, the benefit of the thread pool will be apparent.

The 'overhead' you mention is nothing to do with ExecutorService, it is caused by multiple threads synchronizing on Math.random, creating lock contention.
So yes, you are missing something (and the 'correct' answer below is not actually correct).
Here is some Java 8 code to demonstrate 8 threads running a simple function in which there is no lock contention:
import java.util.ArrayList;
import java.util.List;
import java.util.concurrent.CountDownLatch;
import java.util.concurrent.ExecutorService;
import java.util.concurrent.Executors;
import java.util.concurrent.TimeUnit;
import java.util.function.DoubleFunction;
import com.google.common.base.Stopwatch;
public class ExecServicePerformance {
private static final int repetitions = 120;
private static int totalOperations = 250000;
private static final int cpus = 8;
private static final List<Batch> batches = batches(cpus);
private static DoubleFunction<Double> performanceFunc = (double i) -> {return Math.sin(i * 100000 / Math.PI); };
public static void main( String[] args ) throws InterruptedException {
printExecutionTime("Synchronous", ExecServicePerformance::synchronous);
printExecutionTime("Synchronous batches", ExecServicePerformance::synchronousBatches);
printExecutionTime("Thread per batch", ExecServicePerformance::asynchronousBatches);
printExecutionTime("Executor pool", ExecServicePerformance::executorPool);
}
private static void printExecutionTime(String msg, Runnable f) throws InterruptedException {
long time = 0;
for (int i = 0; i < repetitions; i++) {
Stopwatch stopwatch = Stopwatch.createStarted();
f.run(); //remember, this is a single-threaded synchronous execution since there is no explicit new thread
time += stopwatch.elapsed(TimeUnit.MILLISECONDS);
}
System.out.println(msg + " exec time: " + time);
}
private static void synchronous() {
for ( int i = 0; i < totalOperations; i++ ) {
performanceFunc.apply(i);
}
}
private static void synchronousBatches() {
for ( Batch batch : batches) {
batch.synchronously();
}
}
private static void asynchronousBatches() {
CountDownLatch cb = new CountDownLatch(cpus);
for ( Batch batch : batches) {
Runnable r = () -> { batch.synchronously(); cb.countDown(); };
Thread t = new Thread(r);
t.start();
}
try {
cb.await();
} catch (InterruptedException e) {
throw new RuntimeException(e);
}
}
private static void executorPool() {
final ExecutorService es = Executors.newFixedThreadPool(cpus);
for ( Batch batch : batches ) {
Runnable r = () -> { batch.synchronously(); };
es.submit(r);
}
es.shutdown();
try {
es.awaitTermination( 10, TimeUnit.SECONDS );
} catch (InterruptedException e) {
throw new RuntimeException(e);
}
}
private static List<Batch> batches(final int cpus) {
List<Batch> list = new ArrayList<Batch>();
for ( int i = 0; i < cpus; i++ ) {
list.add( new Batch( totalOperations / cpus ) );
}
System.out.println("Batches: " + list.size());
return list;
}
private static class Batch {
private final int operationsInBatch;
public Batch( final int ops ) {
this.operationsInBatch = ops;
}
public void synchronously() {
for ( int i = 0; i < operationsInBatch; i++ ) {
performanceFunc.apply(i);
}
}
}
}
Result timings for 120 tests of 25k operations (ms):
Synchronous exec time: 9956
Synchronous batches exec time: 9900
Thread per batch exec time: 2176
Executor pool exec time: 1922
Winner: Executor Service.

I don't think this is at all realistic since you're creating a new executor service every time you make the method call. Unless you have very strange requirements that seems unrealistic - typically you'd create the service when your app starts up, and then submit jobs to it.
If you try the benchmarking again but initialise the service as a field, once, outside the timing loop; then you'll see the actual overhead of submitting Runnables to the service vs. running them yourself.
But I don't think you've grasped the point fully - Executors aren't meant to be there for efficiency, they're there to make co-ordinating and handing off work to a thread pool simpler. They will always be less efficient than just invoking Runnable.run() yourself (since at the end of the day the executor service still needs to do this, after doing some extra housekeeping beforehand). It's when you are using them from multiple threads needing asynchronous processing, that they really shine.
Also consider that you're looking at the relative time difference of a basically fixed cost (Executor overhead is the same whether your tasks take 1ms or 1hr to run) compared to a very small variable amount (your trivial runnable). If the executor service takes 5ms extra to run a 1ms task, that's not a very favourable figure. If it takes 5ms extra to run a 5 second task (e.g. a non-trivial SQL query), that's completely negligible and entirely worth it.
So to some extent it depends on your situation - if you have an extremely time-critical section, running lots of small tasks, that don't need to be executed in parallel or asynchronously then you'll get nothing from an Executor. If you're processing heavier tasks in parallel and want to respond asynchronously (e.g. a webapp) then Executors are great.
Whether they are the best choice for you depends on your situation, but really you need to try the tests with realistic representative data. I don't think it would be appropriate to draw any conclusions from the tests you've done unless your tasks really are that trivial (and you don't want to reuse the executor instance...).

Math.random() actually synchronizes on a single Random number generator. Calling Math.random() results in significant contention for the number generator. In fact the more threads you have, the slower it's going to be.
From the Math.random() javadoc:
This method is properly synchronized to allow correct use by more than
one thread. However, if many threads need to generate pseudorandom
numbers at a great rate, it may reduce contention for each thread to
have its own pseudorandom-number generator.

Firstly there's a few issues with the microbenchmark. You do a warm up, which is good. However, it is better to run the test multiple times, which should give a feel as to whether it has really warmed up and the variance of the results. It also tends to be better to do the test of each algorithm in separate runs, otherwise you might cause deoptimisation when an algorithm changes.
The task is very small, although I'm not entirely sure how small. So number of times faster is pretty meaningless. In multithreaded situations, it will touch the same volatile locations so threads could cause really bad performance (use a Random instance per thread). Also a run of 47 milliseconds is a bit short.
Certainly going to another thread for a tiny operation is not going to be fast. Split tasks up into bigger sizes if possible. JDK7 looks as if it will have a fork-join framework, which attempts to support fine tasks from divide and conquer algorithms by preferring to execute tasks on the same thread in order, with larger tasks pulled out by idle threads.

Here are results on my machine (OpenJDK 8 on 64-bit Ubuntu 14.0, Thinkpad W530)
simpleCompuation:6
computationWithObjCreation:5
computationWithObjCreationAndExecutors:33
There's certainly overhead. But remember what these numbers are: milliseconds for 100k iterations. In your case, the overhead was about 4 microseconds per iteration. For me, the overhead was about a quarter of a microsecond.
The overhead is synchronization, internal data structures, and possibly a lack of JIT optimization due to complex code paths (certainly more complex than your for loop).
The tasks that you'd actually want to parallelize would be worth it, despite the quarter microsecond overhead.
FYI, this would be a very bad computation to parallelize. I upped the thread to 8 (the number of cores):
simpleCompuation:5
computationWithObjCreation:6
computationWithObjCreationAndExecutors:38
It didn't make it any faster. This is because Math.random() is synchronized.

The Fixed ThreadPool's ultimate porpose is to reuse already created threads. So the performance gains are seen in the lack of the need to recreate a new thread every time a task is submitted. Hence the stop time must be taken inside the submitted task. Just with in the last statement of the run method.

You need to somehow group execution, in order to submit larger portions of computation to each thread (e.g. build groups based on stock symbol).
I got best results in similar scenarios by using the Disruptor. It has a very low per-job overhead. Still its important to group jobs, naive round robin usually creates many cache misses.
see http://java-is-the-new-c.blogspot.de/2014/01/comparision-of-different-concurrency.html

In case it is useful to others, here are test results with a realistic scenario - use ExecutorService repeatedly until the end of all tasks - on a Samsung Android device.
Simple computation (MS): 102
Use threads (MS): 31049
Use ExecutorService (MS): 257
Code:
ExecutorService executorService = Executors.newFixedThreadPool(1);
int count = 100000;
//Simple computation
Instant instant = Instant.now();
for (int i = 0; i < count; i++) {
double x = Math.random() * Math.random();
}
Duration duration = Duration.between(instant, Instant.now());
Log.d("ExecutorPerformanceTest", "Simple computation (MS): " + duration.toMillis());
//Use threads
instant = Instant.now();
for (int i = 0; i < count; i++) {
new Thread(() -> {
double x = Math.random() * Math.random();
}
).start();
}
duration = Duration.between(instant, Instant.now());
Log.d("ExecutorPerformanceTest", "Use threads (MS): " + duration.toMillis());
//Use ExecutorService
instant = Instant.now();
for (int i = 0; i < count; i++) {
executorService.execute(() -> {
double x = Math.random() * Math.random();
}
);
}
duration = Duration.between(instant, Instant.now());
Log.d("ExecutorPerformanceTest", "Use ExecutorService (MS): " + duration.toMillis());

I've faced a similar problem, but Math.random() was not the issue.
The problem is having many small tasks that take just a few milliseconds to complete. It is not much but a lot of small tasks in series ends up being a lot of time and I needed to parallelize.
So, the solution I found, and it might work for those of you facing this same problem: do not use any of the executor services. Instead create your own long living Threads and feed them tasks.
Here is an example, just as an idea don't try to copy paste it cause it probably won't work as I am using Kotlin and translating to Java in my head. The concept is what's important:
First, the Thread, a Thread that can execute a task and then continue there waiting for the next one:
public class Worker extends Thread {
private Callable task;
private Semaphore semaphore;
private CountDownLatch latch;
public Worker(Semaphore semaphore) {
this.semaphore = semaphore;
}
public void run() {
while (true) {
semaphore.acquire(); // this will block, the while(true) won't go crazy
if (task == null) continue;
task.run();
if (latch != null) latch.countDown();
task = null;
}
}
public void setTask(Callable task) {
this.task = task;
}
public void setCountDownLatch(CountDownLatch latch) {
this.latch = latch;
}
}
There is two things here that need explanation:
the Semaphore: gives you control over how many tasks and when they are executed by this thread
the CountDownLatch: is the way to notify someone else that a task was completed
So this is how you would use this Worker, first just a simple example:
Semaphore semaphore = new Semaphore(0); // initially the semaphore is closed
Worker worker = new Worker(semaphore);
worker.start();
worker.setTask( .. your callable task .. );
semaphore.release(); // this will allow one task to be processed by the worker
Now a more complicated example, with two Threads and waiting for both to complete using the CountDownLatch:
Semaphore semaphore1 = new Semaphore(0);
Worker worker1 = new Worker(semaphore1);
worker1.start();
Semaphore semaphore2 = new Semaphore(0);
Worker worker2 = new Worker(semaphore2);
worker2.start();
// same countdown latch for both workers, with a counter of 2
CountDownLatch countDownLatch = new CountDownLatch(2);
worker1.setCountDownLatch(countDownLatch);
worker2.setCountDownLatch(countDownLatch);
worker1.setTask( .. your callable task .. );
worker2.setTask( .. your callable task .. );
semaphore1.release();
semaphore2.release();
countDownLatch.await(); // this will block until 2 tasks have been completed
And after that code runs you could just add more tasks to the same threads and reuse them. That's the whole point of this, reusing the threads instead of creating new ones.
It is unpolished as f*** but hopefully this gives you an idea. For me this was an improvement compared to no multi threading. And it was much much better than any executor service with any number of threads in the pool by far.