RxJava usage optimization request

Question

Today I tried to solve a small challenge:

You are a big company with 500 offices, you want to compute the global revenue (sum of revenues of each office).

Each office exposes a service to get the revenue. The call takes a certain delay (network, db access, ...).

Obviously, you want global revenue as fast as possible.

Firstly I tried in python with pretty good results:

import asyncio
import time

DELAYS = (475, 500, 375, 100, 250, 125, 150, 225, 200, 425, 275, 350, 450, 325, 400, 300, 175)


class Office:

    def __init__(self, delay, name, revenue):
        self.delay = delay
        self.name = name
        self.revenue = revenue

    async def compute(self):
        await asyncio.sleep(self.delay / 1000)
        print(f'{self.name} finished in {self.delay}ms')
        return self.revenue


async def main(offices, totest):
    computed = sum(await asyncio.gather(*[o.compute() for o in offices]))
    verdict = ['nok', 'ok'][computed == totest]
    print(f'Sum of revenues = {computed} {verdict}')


if __name__ == "__main__":
    offices = [Office(DELAYS[i % len(DELAYS)], f'Office-{i}', 3 * i + 10) for i in range(500)]
    totest = sum(o.revenue for o in offices)
    start = time.perf_counter()
    asyncio.run(main(offices, totest))
    end = time.perf_counter()
    print(f'Ends in {(end-start)*1000:.3f}ms')

On my computer it takes around 500ms, the ideal case (because 500ms is the maximum delay)

Next, I tried in java with RxJava:

import java.util.concurrent.TimeUnit;

public class Office {
    private int sleepTime;
    private String name;
    private int revenue;

    public Office(int sleepTime, String name, int revenue) {
        this.sleepTime = sleepTime;
        this.name = name;
        this.revenue = revenue;
    }

    public int getRevenue() {
        return revenue;
    }

    public int compute() {
        try {
            TimeUnit.MILLISECONDS.sleep(this.sleepTime);
        } catch (InterruptedException e) {
            e.printStackTrace();
        }
        System.out.printf("%s finished in %dms on thread %d%n", this.name, this.sleepTime, Thread.currentThread().getId());
        return this.revenue;
    }
}

import io.reactivex.Flowable;
import io.reactivex.schedulers.Schedulers;

import java.time.Duration;
import java.time.Instant;
import java.util.ArrayList;

public class Tester {
    private static int[] DELAYS = {475, 500, 375, 100, 250, 125, 150, 225, 200, 425, 275, 350, 450, 325, 400, 300, 175};

    public static void main(String[] args) {
        final ArrayList<Office> offices = new ArrayList<>();

        for (int i = 0; i < 500; i++) {
            offices.add(new Office(DELAYS[i % DELAYS.length], String.format("Office-%d", i), 3 * i + 10));
        }
        int totest = offices.stream().mapToInt(Office::getRevenue).sum();

        final Instant start = Instant.now();
        final Flowable<Office> officeObservable = Flowable.fromIterable(offices);
        int computation = officeObservable.parallel(500).runOn(Schedulers.io()).map(Office::compute).reduce(Integer::sum).blockingSingle();
        boolean verdict = computation == totest;
        System.out.println("" + computation + " " + (verdict ? "ok" : "nok"));
        final Instant end = Instant.now();

        System.out.printf("Ends in %dms%n", Duration.between(start, end).toMillis());

    }
}

On my computer, it takes around 1000ms (with a pool of 500 threads !!).

Of course, I tried with different number of threads but results are worst or similar.

I don't want to compare Python and Java, I just want:

Explanations if I did mistakes

A better approach?

Also, python async uses only one thread but in Java I didn't find how not to use multithreading to have a similar result.

Maybe someone could help me? :-)

Does total time decrease when you limit office count to 50? If yes, does it decrease linearly with count or there is sudden increase after some number? — Tuby
– Tuby, Commented Feb 8, 2019 at 19:15

wargre · Accepted Answer · 2019-02-08 19:43:49Z

It quite simple. On the python side, you wait in async mode ( not blocking) on the java side, you wait with a blocking code hence the difference.

The correct code in java should be:

package com.test;

import io.reactivex.Flowable;
import io.reactivex.Single;
import io.reactivex.schedulers.Schedulers;
import org.reactivestreams.Publisher;

import java.time.Duration;
import java.time.Instant;
import java.util.ArrayList;
import java.util.concurrent.TimeUnit;


public class TestReactive {

    public static class Office {
        private int sleepTime;
        private String name;
        private int revenue;

        public Office(int sleepTime, String name, int revenue) {
            this.sleepTime = sleepTime;
            this.name = name;
            this.revenue = revenue;
        }

        public int getRevenue() {
            return revenue;
        }

        public Publisher<Integer> compute() {
            return Single.just("")
                    .delay(this.sleepTime, TimeUnit.MILLISECONDS)
                    .map(x-> {
                        System.out.printf("%s finished in %dms on thread %d%n", this.name, this.sleepTime, Thread.currentThread().getId());
                        return this.revenue;
                    }).toFlowable();
        }
    }

    private static int[] DELAYS = {475, 500, 375, 100, 250, 125, 150, 225, 200, 425, 275, 350, 450, 325, 400, 300, 175};

    public static void main(String[] args) {
        final ArrayList<Office> offices = new ArrayList<>();

        for (int i = 0; i < 500; i++) {
            offices.add(new Office(DELAYS[i % DELAYS.length], String.format("Office-%d", i), 3 * i + 10));
        }
        int totest = offices.stream().mapToInt(Office::getRevenue).sum();

        final Instant start = Instant.now();

        final Flowable<Office> officeObservable = Flowable.fromIterable(offices);
        int computation = officeObservable.parallel(2).runOn(Schedulers.io()).flatMap(Office::compute).reduce(Integer::sum).blockingSingle();
        boolean verdict = computation == totest;
        System.out.println("" + computation + " " + (verdict ? "ok" : "nok"));
        final Instant end = Instant.now();

        System.out.printf("Ends in %dms%n", Duration.between(start, end).toMillis());

    }

}

Edit: I set a parallel of 2 but who care, you can put a single thread as it is not a CPU limit issue here.

Etienne · Accepted Answer · 2019-02-08 22:10:01Z

After a lot of tries (thanks M. T for you help), finally I have a good Java implementation!

public class Office {
    private int sleepTime;
    private int revenue;

    public Office(int sleepTime, int revenue) {
        this.sleepTime = sleepTime;
        this.revenue = revenue;
    }

    public int getRevenue() {
        return revenue;
    }

    public Single<Integer> compute() {
        return Single.timer(sleepTime, TimeUnit.MILLISECONDS).map(l -> this.revenue);
    }
}


public class Tester {
    private static int[] DELAYS = {475, 500, 375, 100, 250, 125, 150, 225, 200, 425, 275, 350, 450, 325, 400, 300, 175};

    public static void main(String[] args) {
        final ArrayList<Office> offices = new ArrayList<>();

        for (int i = 0; i < 1_000_000; i++) {
            offices.add(new Office(DELAYS[i % DELAYS.length], 1));
        }
        int totest = offices.stream().mapToInt(Office::getRevenue).sum();

        final Instant start = Instant.now();
        final Flowable<Office> officeObservable = Flowable.fromIterable(offices);
        int computation = officeObservable.flatMapSingle(Office::compute).reduce(Integer::sum).blockingGet();
        boolean verdict = computation == totest;
        System.out.println("" + computation + " " + (verdict ? "ok" : "nok"));
        final Instant end = Instant.now();

        System.out.printf("Ends in %dms%n", Duration.between(start, end).toMillis());
    }
}

This code is blazing fast! 2s for 1_000_000 offices!

Collectives™ on Stack Overflow

RxJava usage optimization request

2 Answers 2

Comments

Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

2 Answers 2

Comments

Comments

Your Answer

Sign up or log in

Post as a guest

Related