YCSB

Links

• To get here, use https://ycsb.site
• Our project docs
• The original announcement from Yahoo!

Getting Started

1. Download the latest release of YCSB:

   curl -O --location https://github.com/brianfrankcooper/YCSB/releases/download/0.17.0/ycsb-0.17.0.tar.gz
   tar xfvz ycsb-0.17.0.tar.gz
   cd ycsb-0.17.0

2. Set up a database to benchmark. There is a README file under each binding directory.

3. Run YCSB command.

   On Linux:

   bin/ycsb.sh load basic -P workloads/workloada
   bin/ycsb.sh run basic -P workloads/workloada

   On Windows:

   bin/ycsb.bat load basic -P workloads\workloada
   bin/ycsb.bat run basic -P workloads\workloada

Running the ycsb command without any argument will print the usage.

See https://github.com/brianfrankcooper/YCSB/wiki/Running-a-Workload for a detailed documentation on how to run a workload.

See https://github.com/brianfrankcooper/YCSB/wiki/Core-Properties for the list of available workload properties.

Building from source

YCSB requires the use of Maven 3; if you use Maven 2, you may see errors such as these.

To build the full distribution, with all database bindings:

mvn clean package

To build a single database binding:

mvn -pl site.ycsb:mongodb-binding -am clean package

Running multiple instances and latency percentiles

In general, you shall be interested in 99% percentile (P99) of the latency distribution, and the rest of the tail - 99.9%, 99.99%, 99.999%. The difference between the amount of requests that will be observed by a user that fall into 95% (P95) percentile and 99% percentile may be sufficiently large.

For example, see "How Many Nines?" at https://bravenewgeek.com/everything-you-know-about-latency-is-wrong/. The formula to calculate probability of how many clients will observe a specific percentile is:

Probability_to_observe = 1 - Percentile ^ Requests

That is why almost 30% of the users will observe latency worse than P99 just by loading the default google.com web page:

1 - 0.99 ^ 30 = 0.27

Remember, that

• latencies percentiles can't be averaged. Don't fall into this trap. Neither latency averages, nor P99 averages do not make any sense.

If you run multiple loaders dump result histograms with:

-p hdrhistogram.fileoutput=true
-p hdrhistogram.output.path=file.hdr

merge them manually and extract required percentiles out of the joined result.

Remember that running multiple workloads may distort original workloads distributions they were intended to produce.

Merging HDR histogram percentiles

HdrHistogram can serialize its data to HDR files. Use CLI tool to do different operations with your saved histograms https://github.com/nitsanw/HdrLogProcessing.

You shall be interested in 3 functions:

• Union - to combine result histograms
• Summarize - to extract latency percentiles
• An ability to print the result into the CSV file and extract tags

To extract HDR content into CSV file format use from https://github.com/HdrHistogram/HdrHistogram/:

java -cp HdrHistogram-2.1.9.jar org.HdrHistogram.HistogramLogProcessor -i file.hdr -o output_${tag}.csv -csv -tag ${tag}