Hi, @jing feng I think we can do a test with the implementation of netty as an example. There is an IntObjectHashMapBenchmark in netty, @jing feng Can you help me to see the specific implementation and effect?
I certainly agree with what you said, staying the same before the investigation, which is why we did not start submitting pr or modifying the code. Regarding the two scenarios you said, I personally think that we will discuss it after we have finished the test. Thank you very much for your advice on how to do benchmarking. I think you can discuss it with @stormyu here and improve the benchmark. And after that, we will discuss which maps we can replace and need to replace. 徐靖峰 <[email protected]> 于2018年12月19日周三 上午11:19写道： > > LongObjectHashMap and IntObjectHashMap 无论是 netty 的实现还是像 JCtool, hppc > 这样实现，据我所知，并不是为了快而诞生的，他们突出的一个特性是相比 HashMap 节约了内存的使用量，并且可能在解决 Hash 冲突方面做了一些优化， > > > A hash map implementation of LongObjectMap < > https://netty.io/4.1/api/io/netty/util/collection/LongObjectMap.html> > that uses open addressing for keys. To minimize the memory footprint, this > class uses open addressing rather than chaining. Collisions are resolved > using linear probing. Deletions implement compaction, so cost of remove can > approach O(N) for full maps, which makes a small loadFactor recommended. > > 线性探测和链表的方式解决 Hash 冲突其实本质上是一种 trade off，在不同场景下可以说各有优势，但在实际使用场景下，比如 > Dubbo，这两者的差距我觉得很难考量，或者请允许我猜测，他们的差距是微乎其微的。我个人建议是负责测试同学不能仅仅评测单机下的 > micro-benchmark ，而是要在实际场景中测试两者的差距，包括 qps > 和内存使用量，这工作我觉得是很大的。毕竟我们引入了第三方的依赖，需要有足够的数据佐证这样的改动是有意义的，否则，我在 benchmark > 之前，会认为这样的 trade off 是我倾向于保持原样，使用原有的 HashMap。 > > 1. Dubbo 场景中大量的 Map<Long,Object>，这里的 Long 主要是 requestId，在并发请求中，需要快速 put, > get , remove ，应该还是需要 focus 在并发性能方面，而非内存占用，所以 synchronized LongObjectHashMap > 意义不是很大。 > 2. 数据需要长期存在于内存中的场景，LongObjectHashMap 有其存在的意义，例如 dubbo > 的一些服务元数据，方法信息等等，他们不需要频繁变更，却需要长期占据内存，这比较适合使用新增的 Map 数据结构。 > > 我的观点是主要围绕上述两点去展开工作。 > > 另外，如果想要替换 ConcurrentHashMap，可以测评下 JCtools 的并发 Map，据称是比 jdk 更强。 > > Last, 这样的 issue 非常有意义，我会持续跟进 > ------------------------------------ > > English：pls copy the above Chinese issue to google translation ，forgive my > laziness, hhh. > > > 在 2018年12月19日，上午11:04，yuhang xiu <[email protected]> 写道： > > > > Hi, jingfeng, > > > > I think we should have a benchmark about the sync long object map and > > concurrent map. > > Personally think the concurrent map is better, but we should have a > > benchmark about this two cases. > > > > Another thing, maybe my description is problematic, we are going to > replace > > longobjectmap and other maps such as intobjectmap. Not just > longobjectmap. > > There also has some intobjectmap in dubbo. > > After the performance, @stormyu will give us that which map can be > replaced. > > > > And the longobjectmap and intobjectmap has many impls (not just in > netty). > > Could u give some ideas about this? > >
