Benchmarking Redis

2 months ago I started working on an ORM for Redis a key-value database. Now I am about to rewrite parts of it because I didn't test and measure everything beforehand.

This time I decided to measure the performance of the different Redis commands. I want to be sure that any optimizations I might implement really helps.

A Redis database is a big mapping. Its keys are strings and the values can be strings, lists, sets, ordered sets, and hashes. Here I’ll just consider strings and hashes. The 2 types of commands I'm interested in are:

The command prefixed by H are for hashes, the ones without prefix are for regular strings. Commands with an M are the ones that can set or get multiple values. To learn more about how Redis works check out the documentation.

I wrote a little script to test how fast the different commands are depending on the number of items to write or read for each operation. I configured Redis to use the Append-only-file like this:

appendonly yes
appendfsync always

Write operations will surely be faster without it. but that's how I use Redis. I do this because I don’t want to lose data, ever!

Here are the results:

Times are in micro-seconds per items
1 167.25 163.04 216.26 203.46
2 164.99 161.72 112.71 109.85
10 156.37 157.34 35.41 33.73
100 156.26 164.23 22.31 19.31
1000 175.82 161.74 31.10 29.56
1 816.10 851.05 887.86 902.71
2 908.21 939.67 560.31 587.40
10 965.12 983.00 218.81 227.05
100 945.03 1037.13 120.09 110.63
1000 897.69 901.27 106.49 101.37

Hash operations are a little bit faster than the ones on strings, probably because there are fewer keys in a hash than in the full Redis mapping. The difference is not too significant though.

*SET operations are faster than *MSET operations for a single element, and for 2 elements or more *MSET operations are faster. No surprise here. But something worth noting is that *MSET operations are MUCH faster than *SET operations with 2 or more elements. They can be as much as 5 times faster when reading and 9 times faster when writing. Something to keep in mind when you have to update 5000 keys in a single batch.