RFR: 8340547: Starting many threads can delay safepoints [v2]

[Oli Gillespie]

On Fri, 20 Sep 2024 15:56:08 GMT, Oli Gillespie <ogillespie at openjdk.org> wrote:

>> Mitigate the impact of JVM_StartThread on safepoint synchronization, by adding a new ThreadStart_lock which limits the number of JVM_StartThread invocations competing for Threads_lock at any given time to 1.
>> This gives a VM thread trying to call a safepoint a much better chance of acquiring Threads_lock when there are many JVM_StartThread invocations in flight, at the cost of one extra lock/unlock for every new thread.
>> 
>> Can be disabled with new diagnostic flag `-XX:-UseExtraThreadStartLock`.
>> 
>> Before (ThreadStartTtsp.java is shared in JDK-8340547):
>> 
>> java -Xlog:safepoint ThreadStartTtsp.java | grep -o 'Reaching safepoint: [0-9]* ns'
>> Reaching safepoint: 1291591 ns
>> Reaching safepoint: 59962 ns
>> Reaching safepoint: 1958065 ns
>> Reaching safepoint: 14456666258 ns <-- 14 seconds!
>> ...
>> 
>> 
>> After:
>> 
>> java -Xlog:safepoint ThreadStartTtsp.java | grep -o 'Reaching safepoint: [0-9]* ns'
>> Reaching safepoint: 214269 ns
>> Reaching safepoint: 60253 ns
>> Reaching safepoint: 2040680 ns
>> Reaching safepoint: 3089284 ns
>> Reaching safepoint: 2998303 ns
>> Reaching safepoint: 4433713 ns <-- 4.4ms
>> Reaching safepoint: 3368436 ns
>> Reaching safepoint: 2986519 ns
>> Reaching safepoint: 3269102 ns
>> ...
>> 
>> 
>> 
>> **Alternatives**
>> 
>> I considered some other options for mitigating this. For example, could we reduce the time spent holding the lock in StartThread? Most of the time is spent managing the threads list for ThreadSMR support, and each time we add a thread to that list we need to copy the whole list and free every entry in the original, which is slow. But I didn't see an easy way to avoid this.
>> I also looked at some kind of signal from the VM thread that it is ready to start synchronizing that StartThread could check before trying to grab Threads_lock, but I didn't find anything better than this extra lock.
>
> Oli Gillespie has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Fix build and address comments

Oh I see what you mean now, sorry - what penalty does the fix add to the normal case? On my x86 host:


@Benchmark
public Thread startThread() {
    Thread t = new Thread(() -> {});
    t.start();
    return t;
}


With fix: 16285 ± 189 op/s.
Without fix: 16375 ± 158  ops/s

So possibly within the noise, or perhaps somewhere around ~0.5% overhead. It's the cost of acquiring one extra lock, which intuitively is negligible compared to the significant cost of starting a thread with SMR. Starting threads is already slow, adding a little to that to avoid hard to debug vm-wide delays seems sensible to me.

-------------

PR Comment: https://git.openjdk.org/jdk/pull/21111#issuecomment-2367781829