In previous MC versions, we had a rather simple internal scheduler
for delayed tasks that would just keep pushing task back until desired
tick was reached.
The method it called to schedule the task changed behavior in 1.14, and now
this scheduler is not working nowhere near what it was supposed to be doing.
This was causing long delayed task to eat up CPU (In Oversleep for example)
Rewrite this to just use the CraftScheduler for scheduling delayed tasks.
Once this was fixed, it became quite clear the code that delayed ticket
additions for chunks based on distance was clearly not right, as it was
tested on the previous broken logic.
So the ticket delay process has been vastly revamped to be even smarter.
Chunks behind the player can load slower than the chunks in front of the player.
We also can delay ticket adding until one of its neighbors has loaded, as
this lets us get a smoother spiral out for the chunks (minus frustum intent).
Additionally on frustum previous commit inadvertently broke frustum trying to
fix an issue when the real fix lied elsewhere, so restore chunk priority so
it works again.
Massive update to light to improve performance and chunk loading/generation.
1) Massive bit packing/unpacking optimizations and inlining.
A lot of performance has to do with constant packing and unpacking of bits.
We now inline a most bit operations, and re-use base x/y/z bits in many places.
This helps with cpu level processing to just do all the math at once instead
of having to jump in and out of function calls.
This much logic also is likely over the JVM Inline limit for JIT too.
2) Applied a few of JellySquid's Phosphor mod optimizations such as
- ensuring we don't notify neighbor chunks when neighbor chunk doesn't need to be notified
- reduce hasLight checks in initializing light, and prob some more, they are tagged JellySquid where phosphor influence was used.
3) Optimize hot path accesses to getting updating chunk to have less branching
4) Optimize getBlock accesses to have less branching, and less unpacking
5) Have a separate urgent bucket for chunk light tasks. These tasks will always cut in line over non blocking light tasks.
6) Retain chunk priority while light tasks are enqueued. So if a task comes in at high priority but the queue is full
of tasks already at a lower priority, before the task was simply added to the end. Now it can cut in line to the front.
this applies for both urgent and non urgent tasks.
7) Buffer non urgent tasks even if queueUpdate is called multiple times to improve efficiency.
8) Fix NPE risk that crashes server in getting nibble data
Fixes#3489Fixes#3363
Mojang implemented a cache like chunks have, but this cache
is accessed by multiple threads and is totally not safe.
So just remove it
Fixes#3466
Also missed a pooled nibble release, so slid that in there too.
The nibble pooling for NBT Tags was 'semi' leaked from loaded chunks
as we store the NBT Tag of Tile Entities in a Chunk, but don't process
them and remove them until chunk reaches Entity Ticking status....
This caused some phantom references to persist causing high memory use
of these chunks.
So I just got rid of pooling from NBT deserialization and we'll have to
take the hit on memory allocations there because too many cascading concerns
with anyone using NBT Tag Byte Arrays.
Fixes#3431
synchronized arraydeque ends up still being way faster.
Kinda shocked how much that strategy was using, it wasn't really
that complicated... but oh well, this is even simpler and not
seeing blocked threads show up at all in profiling because
the lock is held for such a short amount of time.
also because most uses are on either server thread pool or chunk load pool.
Also optimize the pooling of nibbles to not register Cleaner's
for Light Engine directed usages, as we know we are properly
controlling clean up there, so we don't need to rely on GC.
This will return them to the pool manually, saving a lot of Cleaners.
Closes#3417
Dynmap accessed the raw bytes because it utilized NBT locally, but the
NBTTagcompound was garbage collected while the bytes were still being used.
This will return getBytes() back to being safe, and add a new PoolSafe method
that will prevent the additional allocations for general chunk loading.
Also fixed applyPatches for people with paths in their working directory
if they have mcdev sources built.
Massively reduces memory allocation of 2048 byte buffers by using
an object pool for these.
Uses lots of advanced new capabilities of the Paper codebase :)
Targets 3072 * 8 buffers per 1GB of heap memory up to a max consideration
of 6GB of heap (any more over 6GB won't give more nibble pool)
You can control the 3072 number by setting -DPaper.nibbleBucketSize=2048
Remember this number is * 8 then * heap memory in GB
That is 98304 objects for 4GB of memory, at 2064 bytes roughly, meaning 194MB
You may also control max number of pooled objects directly instead of any
dynamic calculation using -DPaper.maxNibblePoolSize=1024000
While this will use more old generation by a tad bit, allocation rate will drop
significantly, causing less young generation GC's.
This commit has gone through extensive testing for over a day and confident
it no longer has any issues with light corruption.
