apache / apache/arrow-java

[java][Vector][Performance] PromotableWriter.setPosition resolving to itable dispatch on C2 compiled path.

未关闭
#1,238 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看
Type: enhancement
主要语言
Java
星标
94
派生
152
平均合并
3 天 16 小时
30 天内合并 PR
11

描述

### Describe the enhancement requested

This is my first issue report here. Thanks for all the wonderful work.

Usecase: I am planning to replace our parquet-mr based parquet writer with arrow-java to arrow-rs via SharedMemory IPC. Each Kubernetes pod writes about 3 to 4 thousand event per second per pod. The original schema does have some heavy nesting (thrift influence) and schema redesign while desirable but is a much longer exercise. In the process of proof of concept, was able to get the whole flow working but when running some sample shadow flows, saw that all dispatch via
```java
PromotableWriter.setPosition(int index)
```
would result in a itable dispatch. Looking at it, there could be multiple implementations of FieldWriter's at callsite and C2 fails to convert it into monomorphic or bimorphic calls. Was able to get a synthetic benchmark setup inside the performance module and can see the same in JMH runs too.

Image

[This](https://shipilev.net/blog/2015/black-magic-method-dispatch/) wonderful writeup by Aleksey Shipilev speaks about it in great detail.
We can convert the itable dispatch to a vtable dispatch by the method to be dispatched over AbstractFieldWriter. I will do more tests around the same and can submit the PR here. Tried searching over the issues and this doesn't seem to have been reported so far. Just curious if I am doing something wrong when using ListVector's (and hence holding back attaching the PR straight away!).

Just pasting the screen grab of the actual struct writer used in the benchmark. The benchmark is produced by claude (as one can make it out looking at the comments, but it seemed pretty ok to me).

Image

贡献指南

打开贡献指南

调研方向

Start with PromotableWriter.setPosition(int) and the AbstractFieldWriter call path described in the issue. Reproduce the dispatch behavior using the JMH benchmark in the performance module, then verify that the resulting change improves the measured dispatch without regressions.

由索引模型根据 Issue 内容生成。

评估

技术栈
java
领域
performance
Issue 类型
功能
难度
4/5
预计耗时
3-5 天
活跃度
冷清
描述清晰度
基本清楚
新手友好度
45/100

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。