Details
Major Description
It was a bit of a puzzle why we can do a full scan of a disk in about 30 seconds during FSDir() or getVolumeMap(), but the same disk took 11 minutes to do Upgrade replication via hardlinks. It turns out that the org.apache.hadoop.fs.FileUtil.createHardLink() method does an outcall to Runtime.getRuntime().exec(), to utilize native filesystem hardlink capability. So it is forking a full-weight external process, and we call it on each individual file to be replicated.
As a simple check on the possible cost of this approach, I built a Perl test script (under Linux on a production-class datanode). Perl also uses a compiled and optimized p-code engine, and it has both native support for hardlinks and the ability to do "exec".
- A simple script to create 256,000 files in a directory tree organized like the Datanode, took 10 seconds to run.
- Replicating that directory tree using hardlinks, the same way as the Datanode, took 12 seconds using native hardlink support.
- The same replication using outcalls to exec, one per file, took 256 seconds!
- Batching the calls, and doing 'exec' once per directory instead of once per file, took 16 seconds.
Obviously, your mileage will vary based on the number of blocks per volume. A volume with less than about 4000 blocks will have only 65 directories. A volume with more than 4K and less than about 250K blocks will have 4200 directories (more or less). And there are two files per block (the data file and the .meta file). So the average number of files per directory may vary from 2:1 to 500:1. A node with 50K blocks and four volumes will have 25K files per volume, or an average of about 6:1. So this change may be expected to take it down from, say, 12 minutes per volume to 2.