How to process lines of a file in parallel?

Question

I want to read a large file, process each line and insert the results into a database. My goal is to parallelize the processing of the lines, as each process is a longrunning task. Therefore I want one thread to keep reading, multiple threads to keep processing, and one thread keep inserting in chunks to db.

I broke it down as follows:

1) read a file line by line sequentially (easy)

2) send each line to a threadpool (3 threads), as the processing is the long-running task. block further line reading while threadpool is busy.

3) write each processed line from each theadpool to StringBuffer

4) monitor that buffer size, and write the results in chunks to a database (eg each 1000 entries)

ExecutorService executor = Executors.newFixedThreadPool(3);

StringBuffer sb = new StringBuffer();

String line;
AtomicInteger count = new AtomicInteger(0);
while ((line = reader.read()) != null) {
    count.getAndIncrement();
    Future<String> future = executor.submit(() -> {
        return processor.process(line);
    });

    //PROBLEM: this blocks until the future returns
    sb.append(future.get());

    if (count.get() == 100) {
        bufferChunk = sb;
        count = new AtomicInteger(0);
        sb = new StringBuffer();

        databaseService.batchInsert(bufferChunk.toString());
    }
}

Problems:

• future.get() will always block the reader until one future returns a result

• the buffer "monitoring" is probably not done right

Probably I'm not doing this the right way. But how can I achieve this?

Sidenote: filesize is about 10GB, so I cannot first read the entire file into memory to prepare the parallel tasks.

Pulverulent

is that screaming use a bounded queue just for me? :| – Bird

Answer 1

I find the following solution elegant. It is only one of the many possible, but it is conceptually simple and

• it throttles the reads,
• accumulates just the minimum amount of state to report ready at the end
• does not require explicit handling of threads

I am only putting the actual test method here with the complete test setup and auxiliary data structures available in a dedicated GitHub repo:

private final AtomicInteger count = new AtomicInteger();

private final Consumer<String> processor = (value) -> {
    count.incrementAndGet();
};

@Test
public void onlyReadWhenExecutorAvailable() throws Exception {

    Executor executor = Executors.newCachedThreadPool();

    CompletableFuture<Void> done = CompletableFuture.completedFuture(null);
    for (Semaphore semaphore = new Semaphore(CONCURRENCY_LEVEL); ; ) {
        String value = reader.read();
        if (value == null) {
            break;
        }

        semaphore.acquire();

        CompletableFuture<Void> future = CompletableFuture.completedFuture(value)
            .thenAcceptAsync(v -> {
                processor.accept(v);
                semaphore.release();
            }, executor);

        done = done.thenCompose($ -> future);
    }
    done.get();

    assertEquals(ENTRIES, count.get());
}

Bluebell

So there is no ExecutorService that just bocks the main thread from submitting if queue is full? – Pulverulent

Not that I know about, it queues and this is what does not work for you. – Bluebell

@Pulverulent I have updated my answer to be really straightforward and avoid loops for thread blocking. Really like it now – Bluebell

Answer 2

1. Read file size. (File.length() method) and split it to your number of threads.
2. Use RandomAccessFile to search any new line characters previous the you indexes found at @1. https://docs.oracle.com/javase/7/docs/api/java/io/RandomAccessFile.html
3. Send to each thread the new indexes/offsets + RandomAccessFile with read access to each.
4. Subclass InputStream to create a new InputStream on top of RandomAccessFile and start reading.

Reeva

Answer 3

After deeper research, I found the BlockingExecutor presented in this answer comes closest to what I'm trying to achieve:

https://mcmap.net/q/1472980/-how-to-implement-blocking-threadpoolexecutor

It basically extends ThreadPoolExecutor combined with a Semaphore lock.

Pulverulent

Cool, the idea seems to be similar to my last effort :) Happy you found your solution! – Bluebell