Sponge: Performance drop after upgrading from SF 3029 to 3064

[Lemmmy]

I am currently running

• SpongeForge version: 3064
• Java version: 1.8.0_151
• Operating System: Ubuntu 16.04.4 LTS

• Plugins/Mods: Plugins (15): Minecraft, Minecraft Coder Pack, SpongeAPI, SpongeForge, ActiveTime, AutoSpy, Buycraft, Dynmap, GriefPrevention, LuckPerms, Nucleus, SpongePrometheusExporter, SpongeSlack, TabManager, kristpay
  Mods (29): Minecraft, Minecraft Coder Pack, Forge Mod Loader, Minecraft Forge, SpongeAPI, SpongeForge, AutoRegLib, Baubles, BiblioCraft, CC Prometheus, CC: Tweaked, Chisel, Computronics, Dynmap, DynmapBlockScan, Iron Chest, Just Enough Items, MinimalPeripherals, OpenComputers, OpenComputers (Core), Plethora, Plethora Core, Quark, Sampling Performance Profiler, The Weirding Gadget, TickProfiler, WorldEdit, asielib, mercurius_updater

Issue Description

Last night we upgraded our server from SpongeForge 3029 to 3064, and have noticed a stable decline in TPS since. This 24-hour graph shows the TPS before upgrading, and since upgrading:

Here is a snippet of a 1-minute Sampler timing:
And the full NPS: https://crashes.switchcraft.pw/2018-04-30-19-13.nps

Additionally, a timings report: https://timings.aikar.co/?id=42dd995e9fa042ce9db99a8f1a35e9a6

Phit suggests this may be related to #1706.

[ProsperCraft]

If it helps, here is some more data using 3064.

https://timings.aikar.co/?id=934635a70aa549adab990405becbd642

Warm Roast-
sponge3064.zip

Sampler-
samplersponge3064.zip

Tickprofiler-
https://gist.github.com/ProsperCraft/dd334abea4d170a861387460af3a360a

LagGoggles-

[gabizou]

Do you have stacktraces enabled for phase states in your config per chance?

[Lemmmy]

generate-stacktrace-per-phase=false this? It's false in our config

[ProsperCraft]

Timings includes sponge configs right? @gabizou

[ProsperCraft]

generate-stacktrace-per-phase=false

In our config.

[ProsperCraft]

Here are timings from 3059 https://timings.aikar.co/?id=10692105563244718a7f3f8c3836daf3

Sampler nps file from 3059-
sponge3059.zip

[bloodmc]

@phit is correct. The issue is due to commit https://github.com/SpongePowered/SpongeCommon/commit/ba151bbf8b3b83c3a0c00aea7cae4a41d86ca165

This needs to be reverted.

[Lemmmy]

Hey there! Is there any progress on this issue?

[liach]

why did he use array maps?

[SquidDev]

Would it be possible to switch to Int2ObjectOpenHashMaps? They should have substantially better performance characteristics than the ArrayMap, whilst still avoiding the unboxing overhead.

[bloodmc]

@Lemmmy @ProsperCraft Please test latest SF and report back if problem is resolved.

[ProsperCraft]

Looks like you nailed it!

will continue to monitor.

[phit]

closing for now then, if it becomes an issue again reply and I'll reopen!

[ProsperCraft]

Sorry, looks like I spoke too soon.

https://timings.aikar.co/?id=779c21a25943480894bc4481c38e890b

Here is a sampler profile
sponge.zip

Here is a warmroast-
sponge-warmroast.zip

[bloodmc]

@ProsperCraft Please test the following 2 jars

https://www.dropbox.com/s/npevwshdg2xkfxb/spongeforge-1.12.2-2655-7.1.0-BETA-TEST1.jar
https://www.dropbox.com/s/64kxyntig2v0yp3/spongeforge-1.12.2-2655-7.1.0-BETA-TEST2.jar

Let me know if any perform better. If not then the issue has nothing to do with these changes.

[XakepSDK]

Just make samples with 3029 to see diff.

[ProsperCraft]

Test1-
10 players on

Timings- https://timings.aikar.co/?id=0a194a4198074332af6fe12c436997bb
Sampler-
test1.zip