Friday, May 12, 2017

Database IO performance tests

I work with InnoDB and RocksDB storage engines for MySQL and do performance tests to compare both storage engines and storage devices. I have expertise in MySQL and storage engines but not so much in storage devices, so I don't mind running MySQL. Other people have expertise in the layers under MySQL (Linux, storage) and might mind running MySQL. Fortunately, we have benchmark clients for them.

Obviously there is fio and it is my first choiceThe fio team even added support for coordinated omission when I asked for it. Alas it can't generate all of the IO patterns that I need.

It would be great to share a link to io.go here were Domas to publish that code.

I wrote innosim many years ago to simulate InnoDB IO patterns. Docs are here and a helper script to run a sequence of tests is here.

Finally there is db_bench for RocksDB. One challenge with RocksDB is tuning, so I have a script to help with that and use good options to run a sequence of tests in a special pattern. Well, it works as long as I keep the script current and I just updated it today. It runs these benchmarks in order:
  1. fillseq - Put N key-value pairs in key order
  2. overwrite - Put N key-value pairs in random order. Queries done after a key-order load avoid a few sources of overhead that usually should not be avoided, so this shuffles the database.
  3. overwrite - Put key-values pairs in random order. Runs for K seconds.
  4. updaterandom - do read-modify-write in random order. Runs for K of seconds.
  5. readwhilewriting - 1 rate-limited writer and T threads doing Get.
  6. seekrandomwhilewriting - 1 rate-limited writer and T threads doing range scans.

No comments:

Post a Comment