=?utf-8?Q?=E5=9B=9E=E5=A4=8D=EF=BC=9AFiber_?=try out feedback

[Ron Pressler]

Thank you.

I’ve just pushed a fix to the bug you’ve found.

Also note that we’ve recently pushed some changes with a possibly inverse effect to performance.
Again — this is still too soon for benchmarks or drawing any conclusions about speed or footprint. The time for that will come.

Ron


On July 19, 2019 at 4:34:44 PM, Lei Yu (lei.yul at alibaba-inc.com) wrote:

Here is the test case

/*
* @test
* @library /lib/testlibrary
* @summary Test fibers pass token to each others
* @run main PassTokenTest
*/
import java.util.concurrent.CountDownLatch;
import java.util.concurrent.ExecutorService;
import java.util.concurrent.Executors;
import java.util.concurrent.TimeUnit;
import java.util.concurrent.locks.Condition;
import java.util.concurrent.locks.Lock;
import java.util.concurrent.locks.ReentrantLock;
import java.util.function.Consumer;
import static jdk.testlibrary.Asserts.assertTrue;
public class PassTokenTest {
static final int N = 1_000_000;
Runner[] runners = new Runner[16];
public static void main(String[] args) throws Exception {
while (true) {
new PassTokenTest().doTest(FiberScope.background()::schedule);
}
}
private void doTest(Consumer<Runnable> run) throws Exception {
for (int i = 0; i < runners.length; i++) {
runners[i] = new Runner(i);
}
CountDownLatch finishLatch = new CountDownLatch(runners.length);
for (Runner runner : runners) {
run.accept(() -> {
runner.run();
finishLatch.countDown();
});
}
assertTrue(finishLatch.await(60, TimeUnit.SECONDS));
}
final long start = System.currentTimeMillis();
int counter = 0;
final Lock lock = new ReentrantLock();
final Condition cond = lock.newCondition();
class Runner implements Runnable {
private final int ord;
public Runner(int ord) {
this.ord = ord;
}
@Override
public void run() {
while (counter < N) {
lock.lock();
try {
while (counter % runners.length != ord) {
try {
cond.await();
} catch (InterruptedException e) {
e.printStackTrace();
}
}
if (++counter % 10000 == 0) // pass the token
System.out.println(Thread.currentThread().getName() + "\t" + counter +
"\t elapsed=" + (System.currentTimeMillis() - start));
cond.signalAll();
} finally {
lock.unlock();
}
}
}
}}

Lei Yu


------------------------------------------------------------------
发件人：Alan Bateman <Alan.Bateman at oracle.com>
发送时间：2019年7月19日(星期五) 19:05
收件人：郁磊(梁希) <lei.yul at alibaba-inc.com>; loom-dev <loom-dev at openjdk.java.net>
主　题：Re: Fiber try out feedback

On 19/07/2019 09:48, Lei Yu wrote:
> Hi,
>
> Here are some feedback on our try out of loom from Alibaba.
> We have written a test of the message passing between the Fibers to verify the performance under high concurrency.
> 1. An assertion error occurred while running the test(can be reproduced by the code in the attachment):
> java.lang.AssertionError
> at java.base/java.lang.Continuation.resizeStack(Continuation.java:685)
> at java.base/java.lang.Continuation.maybeShrink(Continuation.java:671)
> at java.base/java.lang.Continuation.postYieldCleanup(Continuation.java:465)
> at java.base/java.lang.Continuation.run(Continuation.java:341)
> at java.base/java.lang.Fiber.runContinuation(Fiber.java:367)
> at java.base/java.util.concurrent.ForkJoinTask$RunnableExecuteAction.exec(ForkJoinTask.java:1425)
> ...
>
> I think the deeper reason is that the value of Continuation.sp has been changed between maybeShrink() and resizeStack().
>
> 2. Thaw/Freeze interpreter frame will be slowed down by compute_oop_map significantly.
> [image.png]
>
> If a Fiber is executing as a consumer of a BlockingQueue in loop, we can frequently observe the interpreter frame exists even though compilation occurs,
> thaw/freeze interpreter frame may become a performance bottleneck.
Thanks for the bug report/data. I think the attachments (test case and  
PNG) were dropped by the mail system. Can you re-send with the test case  
inlined or if you have someone with write access to cr.openjdk.java.net  
then maybe they could be published there?

-Alan