Document open source YCSB single node

[robert-s-lee]

Examples below uses _sysb which is defined as a bash function available at the bottom of this doc.

• function test
• concurrency test of one workload
• concurrency test of multiple workloads

Couple of notes:

• Open ssh session
• Open CockroachDB admin UI
• YCSB has workloads a,b,c,d,e,f. Workloads D and E during the test run making them not entrant. The recommended sequence load a, run a,b,c,f,d load e, run e
• insertorder=ordered as CockroachDB's sharding does not require hashed primary key.
• The performance listed here is for illustration purposes only. The actual performance is dependent on hardware and database setup and other factors.
• workload generator and database are running on the same hardware. This will sometimes prevent prevent the workload generator from pushing the database as the CPUs is starved.
• use the latest (0.16 snapshot) YCSB which has support multi-row insert

functional-test

• Start a local CockroachDB cluster
• run one iteration

cockroach sql --insecure  -e "drop table if exists usertable"
_ycsb load a
_ycsb run a

• take a look at CockroachDB Admin UI for workload and hardware metrics. Hardware metrics are used to identify resource constraints that prevents better performance.

• the first bump _ycsb load a. Note each insert statement contains 500 rows.
• the second bump is _ycsb run a

concurrency-test-one-workload

• run the test by stepping the concurrency from 1,2,4,8,16,32,64 and 128

for t in 1 2 4 8 16 32 64 128; do 
  _ycsb_threads=$t; _ycsb run b
  sleep 10 
done

concurrency-test-multiple-workloads

for t in 1 2 4 8 16 32 64 128; do 
  _ycsb_threads=$t
  _ycsb run a &
  _ycsb run b &
  _ycsb run c &
  _ycsb run f &
  wait
  sleep 10
done

ycsb.sh.zip

[jseldess]

@robert-s-lee, does this tie into the larger project to show how to use cockroach workloads to calibrate a cluster? #4303

Or is this something distinct?

[robert-s-lee]

#4220 is the master issue to document YCSB and sysbench. Two purposes: one is to document YCSB ad sysbench, the other is to step through the methodology of starting with a functional test, then adding more concurrency gradually before going too big.

[jseldess]

@johnrk, @mwang1026, @rmloveland, closing this for now. If this topics comes out of your Product Areas' problems to solve or priorities in the future, we can create a new issue with more context.