Migration from MongoDB to TokuMX – Big performance gain
Benchmark
About a month ago I’ve done some benchmarking. I’ve compared TokuMX, TokuMXse, MongoDB 2.6, MongoDB 3.0 and WiredTiger. My use case is mainly based on constant selects and updates, all at the same time. Inserts were made before actual select/update test. Here are my results:
As you can see, TokuMX performed best. This benchmark was with 128MB physical ram on server and 250MB of data being inserted (2 mio rows and 100k select/update queries at the same time in total). Idea was to have larger data set selected/updated, than our actual RAM size is, because our production database is 160GB size and we only have about 80GB of ram. Idea is not to have everything in RAM. (I disabled compression feature of TokuMX in this case, because I didn’t want it to compress data to a point, where it could fit actual RAM size).
Upgrade TokuMX on production
After the benchmark, we decided to gradually upgrade all our MongoDB instances to Community TokuMX release. Recently we came to our Europe datacenter, where TokuMX really showed some great performance. Here is our execution time of one event on production:
psvc-write is call to MongoDB/TokuMX. You can clearly see, when we switched from MongoDB to TokuMX. Previous response time for psvc-write was around 20ms on average and was jumping for quite a lot. Current response time is 2ms and constant. It does not jump anymore and it’s about 10x faster than it used to be. There are several thousands of queries per second on this service, so you can imagine general performance boost for our adserving.
Disk usage and it’s performance impact
Thing to mention at the same time, is disk size. Our database used to be 160GB and after migration to TokuMX, it shrinked to 8GB in total. (At first I thought I lost some data, because this % of compression was incredible, but to my surprise, after manual check, all data was there.) Thanks to this fact, our performance gain was even bigger than it would have been if we didn’t use compression. You can imagine our database was way larger than physical RAM of server, but now 8GB can fit into our RAM without a problem.
Final note
This is not benchmark to promote TokuMX. Idea is to encourage you, to do same benchmark of your infrastructure. Think of your production environment, take actual data from production, create benchmarks based on your use case and you can make some wise choices on improving your current infrastructure.
