Bowen Ding

OK I see metadata is cached at the client side. There should be no master-client rpc overhead then. How is the test data written to Alluxio? Is it fully cached...

@JianyuLi @LKR1996 thanks for the porfiling report.  Looks like it spends quite some time acquiring the block lock when reading the block file. Can you try increasing the number...

 Looks like it's actually the `mSharedMapsLock` that is heavily contended, rather than the individual block locks. Can you take a jstack snapshot and see if the threads are all...

I have a PR that removes the shared lock #16078 . Can you give it a try? @JianyuLi

> @dbw9580 your improvement in #16078 solved the problem issue we met here and in #16040. We are looking forward to this MR be merged in master, thanks! Thanks for...

It's possible, but hard to tell without a detailed description of how you did your tests, and some flamegraph showing what master is spending its time doing. Best regards, Bowen...

> > An experimental run on my laptop. I wonder why `monoBlockStoreReadLocal` is significantly slower than any other ones?

https://github.com/Alluxio/alluxio/blob/8f74e6584b6d2d71be3a3737057808b6f77c44ae/core/common/src/main/java/alluxio/worker/block/io/LocalFileBlockReader.java#L100 javadoc on `java.nio.channels.FileChannel#map`says > For most operating systems, mapping a file into memory is more expensive than reading or writing a few tens of kilobytes of data via the...

I take a closer look at the mmap performance on Linux.  Looks like there is quite a significant difference in perf between Linux and Mac. My microbench is here...

@beinan can you please take a look?