heap | Looking into performance of heaps

According to the java doc of PriorityQueue(PriorityQueue)

Creates a PriorityQueue containing the elements in the specified priority queue. This priority queue will be ordered according to the same ordering as the given priority queue.

So we can extend PriorityQueue as CustomComparatorPriorityQueue to hold the desired comparator and the Collection we need to heapify. Then call new PriorityQueue(PriorityQueue) with an instance of CustomComparatorPriorityQueue.

Below is tested to work in Java 15.

No, not until closure HRTB inference is fixed. Current workarounds include using function pointers instead or implementing a helper trait on custom structs -- the helper trait is needed regardless of approach until higher-kinded types are introduced in Rust.

Playground

To avoid returning a Box, you would need the type parameter I to be generic over the lifetime 'a, so that you can use it with any lifetime (in a for<'a> bound, for example). Unfortunately, as discussed in a similar question, Rust does not yet support higher-kinded types (type parameters that are themselves generic over other type parameters), so we must use a helper trait:

Unfortunately (?), the easiest way to know for sure what those categories map to is to look at OpenJDK source code. The NMT tag you are looking for is mtServiceability. This would show that "serviceability" are basically diagnostic interfaces in JDK/JVM: JVMTI, heap dumps, etc.

But the same kind of thing is clear from observing that stack trace sample you are showing mentions ThreadStackTrace::dump_stack_at_safepoint -- that is something that dumps the thread information, for example for jstack, heap dump, etc. If you have a suspicion for the memory leak in that code, you might try to build a MCVE demonstrating it, and submitting the bug against OpenJDK, or showing it to a fellow OpenJDK developer. You probably know better what your application is doing to cause thread dumps, focus there.

That being said, I don't see any obvious memory leaks in StackFrameInfo, neither can I reproduce any leak with stress tests, so maybe what you are seeing is "just" thread dumping over the larger and larger thread stacks. Or you capture it when thread dump is happening. Or... It is hard to say without the MCVE.

Update: After playing with MCVE, I realized that it reproduces with 17.0.1, but not with either mainline development JDK, or JDK 18 EA, or JDK 17.0.2 EA. I tested with 17.0.2 EA before, so was not seeing it, dang. Bisection between 17.0.1 and 17.0.2 EA shows it was fixed with JDK-8273902 backport. 17.0.2 releases this week, so the bug should disappear after you upgrade.

V8 developer here.

are small integers pooled or not?

They are not (at least not right now). That said, this is a small implementation detail and could be done either way: it would totally be possible to use the constant pool for Smis. I suppose the decision to build special machinery for Smis (instead of reusing the general-purpose constant pool) was made because things turned out to be more efficient that way.

it is not worth it pooling small integers if sizeof(int) < sizeof(int *)

The details are different (a Smi is not an int, and constant pool slots are referenced by index rather than C++ pointer), but this reasoning does go in the right direction: avoiding indirections can save time and memory.

are smis also allocated on the heap?

Yes, everything is allocated on the heap. The stack is only useful for temporary (and sufficiently small) things; that's largely unrelated to the type of thing.

The "trick" of Smis is that they're not stored as separate objects: when you have an object that refers to a Smi, such as let foo = {smi: 42}, then the value 42 can be smi-encoded and stored directly inside the "foo" object (whereas if the value was 42.5, then the object would store a pointer to a separate "HeapNumber"). But since the object is on the heap, so is the Smi.

@DanielCruz

What I understand [...] is that constant small integers are pooled. Variable small integers are not.

Nope. Any literal that occurs in source code is "constant". Whether you use let or const for your variables has nothing to do with this.

Allocation elision is an optimization that is outside of and in addition to the as-if rule. Another optimization with the same properties is copy elision (not to be confused with mandatory elision, since C++17): Is it legal to elide a non-trivial copy/move constructor in initialization?.

1. Despite the test, this implementation of Dijkstra may put Ω(E) items in the priority queue. This will cost Ω(E log E) with every comparison-based priority queue.

2. Why not E log V? Well, assuming a connected, simple, nontrivial graph, we have Θ(E log V) = Θ(E log E) since log (V−1) ≤ log E < log V² = 2 log V.

3. The O(E + V log V)-time implementations of Dijkstra's algorithm depend on a(n amortized) constant-time DecreaseKey operation, avoiding multiple entries for an individual vertex. The implementation in this question will likely be faster in practice on sparse graphs, however.

There appears to be something very wrong with the calculated GC CPU time. It's 41010 secs compared to 2737 sec elapsed, which doesn't make sense if you're only running on three capabilities.

This miscalculation means that the calculated MUT CPU time, which is just total CPU time minus INIT, GC, and EXIT time, is actually a large negative number (5073-41010-2 = -35939). This gives a productivity of -35939/5073=-708%. When the MUT seconds are displayed, negative numbers are truncated at zero, to avoid reporting small negative numbers when MUT is very low and there's a clock precision error, which is why the displayed MUT time is 0 instead of -35939.

I don't know why the GC time is so badly miscalculated. My best guess is this. If you're running on Windows, there are known issues with CPU time clock precision, and it's possible that certain unusual patterns of garbage collection timing might result in precision errors occuring in only one direction, slightly overestimating the actual GC time more often than it underestimates it. Over 2.4 million collections (see your GC stats), this difference could accumulate to a huge positive error.

I looked through GitLab issues, and except for the report on general Windows CPU time imprecision and a couple of probably unrelated negative MUT reports here and here, I didn't see anything helpful.

I found this bug too. Workaround is to use incognito window for degugging.

update: Version 95.0.4638.54 has fixed this issue. But in case similar issue come out again, you can always try to avoid it in incognito window.

I believe what happens is that the memory manager tries to align the chunks at the next available 1MB boundary. But as the 1MB arrays actually take up slightly more than 1MB (for storing length and something else), they get arranged with a gap of almost 1MB between them. When reducing the block size by 16 bytes, they suddenly use up the whole memory again.