Performance of Cassandra
See original GitHub issueHi, we made several tests with YCSB comparing mongodb and cassandra in our current setup. At first we wanted to compare just mongodb on red hat enterprise linux 6/7. But also added testing wired tiger and cassandra.
Cassandra is a pretty new technology for us that we are currently testing. In general we had the following settings:
recordcount=6000000
operationcount=100000000
threads=100
fieldlength=250
measurementtype=timeseries
core_workload_insertion_retry_limit=10000
We also tested with a record count of 2.000.000 and 3.000.000. We used workload a, b and f.
All machines there run on identical nodes. We tested virtual servers (like our current infrastructure) and local hard drives on a dedicated machine.
For cassandra we sometimes had timneout during the load phase. That is why we added the “core_workload_insertion_retry_limit=10000” option.
We also used different amount of nodes in our tests. We tested:
- a single mongodb node
- three mongodb nodes
- 1 cassandra node
- 3 cassandra nodes
- 4 cassandra nodes
- 5 cassandra nodes
On cassandra nodes we tested replication factor 1 and 2.
In all our tests cassandra was much slower than mongodb. In general we did not expect this, especially if we increase the nodes in cassandra. But in our case it does not seem to be the case. We already added a few general performance tips. But due to our infrastructure had not splitted the commit and data directory to separate disks.
We repeated most of the tests but in the end concentrated on workload a. Cause the performance difference was on every workload.
In the web we found tests with YCSB there cassandra is performing quite good. So in general our question would be: Are we doing something wrong?
Thanks
Regards Mike
Issue Analytics
- State:
- Created 8 years ago
- Comments:23 (6 by maintainers)
Top GitHub Comments
Cassandra’s DB connection object supports pooled connections and that size can be controlled with the “cassandra.coreconnections” and “cassandra.maxconnections” properties. You might want to check your values and correlate them to your thread size.
closing as stale.