Hi, I did not find any concrete figures or numbers. But as the objects are reused and recycled the initial offset is lost quickly and no java objects are created for new buffers. Thus, in sitatutions with many small io operations you take a lot of pressure from the GC.
Julian Am 07.11.19, 09:26 schrieb "Xiangdong Huang" <[email protected]>: Hi, I did not. But I have tried calling Snappy.compress() and decompress() with on-heap and off-heap bytes about 6 months ago. off-heap compress is very fast. But in that experiment, creating off-heap bytes is time-consuming (because in that version, creating new byte array or bytebuffer is very frequent). I am also interested in the performance of Netty's ByteBuf. Best, ----------------------------------- Xiangdong Huang School of Software, Tsinghua University 黄向东 清华大学 软件学院 Julian Feinauer <[email protected]> 于2019年11月7日周四 下午3:13写道： > Hi folks, > > in iotdb we have a quite frequent use of Javas native ByteBuffer’s for our > IO. I was wondering if anybody did an analysis on object creation / > destruction and so on. > As some may know, the netty project provides an interface called ByteBuf ( > https://netty.io/4.0/api/io/netty/buffer/ByteBuf.html). Then, > implementations based on On-Heap or Off-Heap memory can be used for > specific use cases. Some operations are even zero-copy. > > Furthermore, ByteBuf are Reference counted and recycled, so the > performance will definetly be superior. > The only important question is if anybody already analyzed how much > potential there is and if its worth the effort? > > Julian >
