Hi Robbin,
On 2/10/2019 7:58 pm, Robbin Ehn wrote:
Hi David,
What if the table is full and must be grown?
The table uses chaining, it just means load factor tip over what is
considered a good backing array size.
Coleen raised a good question in a separate discussion, which made me
realize that once the table has been initially populated all subsequent
additions, and hence all subsequent calls to grow() always happen with
the Threads_lock held. So we can't just defer the grow().
That aside, I'd like to know how expensive it is to grow this table.
What are we talking about here?
We use global counter which on write_synchronize must scan all
threads to make sure they have seen the update (there some
optimization to avoid it if there is no readers at all). Since this
table contains the threads, we get double penalized, for each new
thread the synchronization cost increase AND the number of items.
With concurrent reads you still need many thousands of threads, but
I think I saw someone mentioning 100k threads, assuming concurrent
queries the resize can take hundreds of ms to finish. Note that reads
and inserts still in operate roughly at the same speed while
resizing. So a longer resize is only problematic if we do not
respect safepoints.
I think if anything were capable of running 100K threads we would be
hitting far worse scalability bottlenecks than this. But this does seem
problematic.
Thanks,
David
-----
Thanks, Robbin
David
/Robbin
On 2019-10-02 08:46, David Holmes wrote:
Hi Daniil,
On 2/10/2019 4:13 pm, Daniil Titov wrote:
Please review a change that fixes the issue. The problem here is
that that the thread is added to the ThreadIdTable (introduced in
[3]) while the Threads_lock is held by
JVM_StartThread. When new thread is added to the thread table the
table checks if its load factor is greater than required and if so
it grows itself while polling for safepoints.
After changes [4] an attempt to block the thread while holding the
Threads_lock results in assertion in
Thread::check_possible_safepoint().
The fix proposed by David Holmes ( thank you, David!) is to skip
the ThreadBlockInVM inside ThreadIdTable::grow() method if the
current thread owns the Threads_lock.
Sorry but looking at the fix in context now I think it would be
better to do this:
while (gt.do_task(jt)) {
if (Threads_lock->owner() == jt) {
gt.pause(jt);
ThreadBlockInVM tbivm(jt);
gt.cont(jt);
}
}
This way we don't waste time with the pause/cont when there's no
safepoint pause going to happen - and the owner() check is quicker
than owned_by_self(). That partially addresses a general concern I
have about how long it may take to grow the table, as we are
deferring safepoints until it is complete in this JVM_StartThread
usecase.
In the test you don't need all of:
32 * @run clean ThreadStartTest
33 * @run build ThreadStartTest
34 * @run main ThreadStartTest
just the last @run suffices to build and run the test.
Thanks,
David
-----
Testing : Mach 5 tier1 and tier2 completed successfully, tier3 is
in progress.
[1] Webrev: http://cr.openjdk.java.net/~dtitov/8231666/webrev.01/
[2] Bug: https://bugs.openjdk.java.net/browse/JDK-8231666
[3] https://bugs.openjdk.java.net/browse/JDK-8185005
[4] https://bugs.openjdk.java.net/browse/JDK-8184732
Best regards,
Danill
