Benchmarking

Basho Bench is a benchmarking tool created to conduct accurate and repeatable performance tests and stress tests, and to produce performance graphs.

Basho Bench exposes a pluggable driver interface and has been extended to serve as a benchmarking tool against a variety of projects. New drivers can be written in Erlang and are generally less than 200 lines of code.

Installation

You will need:

  1. One or more load-generating machines on which to install basho_bench. Especially when testing larger clusters, a single machine cannot generate enough load to properly exercise the cluster. Do not run the basho_bench instances on the Riak nodes themselves, since the load generation will compete with Riak for resources.
  2. The R statistics language must be installed (somewhere available to you) if you wish to generate graphs (see the Generating Benchmark Graphs section, below).

Download basho_bench

You can download the pre-built packages below, or build it from source.

Building from Source

Prerequisites

  • Erlang must be installed. See Installing Erlang for instructions and versioning requirements. Note: Unless you’re an experienced Erlang developer, we recommend that you use Ubuntu 14.04 LTS (and not CentOS), when building basho_bench from source. Later versions of CentOS (6 and 7) have difficulty with installing and enabling certain parts of the erlang-crypto package, which is required by basho_bench.
  • Install git (to check out the basho_bench code)

Compiling

git clone git://github.com/basho/basho_bench.git
cd basho_bench
make

Usage

Run the basho_bench script, pass in the config file and the directory to generate the results into:

basho_bench --results-dir <results dir> <config file>

If you’ve installed basho_bench from a pre-built package, you must specify full paths for the test results directory and config file. (Also, don’t use the common ~/ shell notation, specify the user’s home directory explicitly)

basho_bench --results-dir /home/username/bench_results/ /etc/basho_bench/riakc_pb.config

The example above will generate results in


If you built ```basho_bench``` from source, you can get away with
relative paths (and the results directory will be created in the
current directory):

```bash
./basho_bench myconfig.config

This will generate results in tests/current/. You will need to create a configuration file. The recommended approach is to start from a file in the examples directory and modify settings using the Configuration section below for reference.

Generating Benchmark Graphs

The output of from running the basho_bench script can be used to create graphs showing the following:

  • Throughput - Operations per second over the duration of the test.
  • Latency at 99th percentile, 99.9th percentile and max latency for the selected operations.
  • Median latency, mean latency, and 95th percentile latency for the selected operations.

Prerequisites

The R statistics language is needed to generate graphs. Note: If necessary, R can be installed on a different machine than the one running basho_bench, and the performance data can be copied (via rsync, for example) from the load testing machine to the one that will be generating and viewing the graphs (such as a desktop).

Installing R on Ubuntu

sudo apt-get install r-base

Installing R on Other Platforms

Follow the instructions for your platform to install R.

Generating Graphs

If you have installed basho_bench from a pre-built package, and you also have R installed on the same machine, you can generate the current result graph with the following:

Rscript --vanilla /usr/lib/basho_bench/lib/basho_bench*/priv/summary.r -i /home/username/bench_results/current/

This will create a results file in


If you have built ```basho_bench``` from source, you can just use
```make```.  To generate a benchmark graph against the current
results, run:

```bash
make results

This will create a results file in tests/current/summary.png.

You can also run this manually:

priv/summary.r -i tests/current

Troubleshooting Graph Generation

For additional help, see the Troubleshooting Graph Generation section of the basho_bench/README.

How does it work?

When Basho Bench starts (basho_bench.erl), it reads the configuration (basho_bench_config.erl), creates a new results directory, and then sets up the test (basho_bench_app.erl and basho_bench_sup.erl).

During test setup, Basho Bench creates the following:

  • One stats process (basho_bench_stats.erl). This process receives notifications when an operation completes, plus the elapsed time of the operation, and stores it in a histogram. At regular intervals, the histograms are dumped to summary.csv as well as operation-specific latency CSVs (e.g. put_latencies.csv for the PUT operation).
  • N workers, where N is specified by the concurrent configuration setting (basho_bench_worker.erl). The worker process wraps a driver module, specified by the driver configuration setting. The driver is randomly invoked using the distribution of operations as specified by the operations configuration setting. The rate at which the driver invokes operations is governed by the mode setting.

Once these processes have been created and initialized, Basho Bench sends a run command to all worker processes, causing them to begin the test. Each worker is initialized with a common seed value for random number generation to ensure that the generated workload is reproducible at a later date.

During the test, the workers repeatedly call driver:run/4, passing in the next operation to run, a keygen function, a valuegen function, and the last state of the driver. The worker process times the operation, and reports this to the stats process when the operation has completed.

Finally, once the test has been run for the duration specified in the config file, all workers and stats processes are terminated and the benchmark ends. The measured latency and throughput of the test can be found in ./tests/current/. Previous results are in timestamped directories of the form ./tests/YYYYMMDD-HHMMSS/.

Configuration

Basho Bench ships with a number of sample configuration files, available in the /examples directory.

Global Config Settings

mode

The mode setting controls the rate at which workers invoke the {driver:run/4} function with a new operation. There are two possible values:

  • `{max}) - generate as many ops per second as possible
  • `{rate, N}) - generate N ops per second, with exponentially distributed interarrival times

Note that this setting is applied to each driver independently. For example, if {rate, 5} is used with 3 concurrent workers, Basho Bench will be generating 15 (i.e. 5 * 3) operations per second.

% Run at max, i.e.: as quickly as possible
{mode, max}

% Run 15 operations per second per worker
{mode, {rate, 15}}

concurrent

The number of concurrent worker processes. The default is 3 worker processes. This determines the number of concurrent clients running requests on API under test.

% Run 10 concurrent processes
{concurrent, 10}

duration

The duration of the test, in minutes. The default is 5 minutes.

% Run the test for one hour
{duration, 60}

operations

The possible operations that the driver will run, plus their “weight,” or likelihood of being run. The default is [{get,4},{put,4},{delete, 1}], which means that out of every 9 operations, GET will be called four times, PUT will be called four times, and DELETE will be called once, on average.

{operations, [{get, 4}, {put, 1}]}.

Operations are defined on a per-driver basis. Not all drivers will implement the GET/PUT operations discussed above. Consult the driver source to determine the valid operations. If you’re testing the HTTP interface, for example, the corresponding operations are GET and UPDATE, respectively.

If a driver does not support a specified operation (asdfput in this example), you may see errors like this:

DEBUG:Driver basho_bench_driver_null crashed: {function_clause,
                                          [{{{basho_bench_driver_null,run,
                                            [asdfput,
                                             #Fun<basho_bench_keygen.4.4674>,
                                             #Fun<basho_bench_valgen.0.1334>,
                                             undefined]}}},
                                           {{{basho_bench_worker,
                                            worker_next_op,1}}},
                                           {{{basho_bench_worker,
                                            max_worker_run_loop,1}}}]}

driver

The module name of the driver that Basho Bench will use to generate load. A driver may simply invoke code in-process (such as when measuring the performance of DETS) or may open network connections and generate load on a remote system (such as when testing a Riak server/cluster).

Available drivers include:

  • `basho_bench_driver_http_raw) - Uses Riak’s HTTP interface to get/update/insert data on a Riak server
  • `basho_bench_driver_riakc_pb) - Uses Riak’s Protocol Buffers interface to get/put/update/delete data on a Riak serve
  • `basho_bench_driver_riakclient) - Uses Riak’s Distributed Erlang interface to get/put/update/delete data on a Riak server
  • `basho_bench_driver_bitcask) - Directly invokes the Bitcask API
  • `basho_bench_driver_dets) - Directly invokes the DETS API

On invocation of the driver:run/4 method, the driver may return one of the following results:

  • `{ok, NewState}) - operation completed successfully
  • `{error, Reason, NewState}) - operation failed but the driver can continue processing (i.e. recoverable error)
  • `{stop, Reason}) - operation failed; driver can’t/won’t continue processing
  • `{‘EXIT’, Reason}) - operation failed; driver crashed

code_paths

Some drivers need additional Erlang code in order to run. Specify the paths to this code using the code_paths configuration setting.

key_generator

The generator function to use for creating keys. Generators are defined in basho_bench_keygen.erl. Available generators include:

  • {sequential_int, MaxKey}) - generates integers from 0..MaxKey` in order and then stops the system. Note that each instance of this keygen is specific to a worker.
  • {partitioned_sequential_int, MaxKey}) - the same as {sequential_int}`, but splits the keyspace evenly among the worker processes. This is useful for pre-loading a large dataset.
  • {partitioned_sequential_int, StartKey, NumKeys}) - the same as partitioned_sequential_int, but starting at the defined StartKeyand going up toStartKey + NumKeys`.
  • {uniform_int, MaxKey}) - selects an integer from uniform distribution of 0..MaxKey`, i.e. all integers are equally probable.
  • {pareto_int, MaxKey}) - selects an integer from a Pareto distribution, such that 20% of the available keys get selected 80% of the time. Note that the current implementation of this generator _may_ yield values larger thanMaxKey` due to the mathematical properties of the Pareto distribution.
  • {truncated_pareto_int, MaxKey}) - the same as{pareto_int}, but will _not> yield values aboveMaxKey`.
  • {function, Module, Function, Args}) - specifies an external function that should return a key generator function. The worker Idwill be prepended toArgs` when the function is called.
  • {int_to_bin, Generator}) - takes any of the above_int` generators and converts the number to a 32-bit binary. This is needed for some drivers that require a binary key.
  • {int_to_str, Generator}) - takes any of the above_int` generators and converts the number to a string. This is needed for some drivers that require a string key.

The default key generator is {uniform_int, 100000}.

Examples:

% Use a randomly selected integer between 1 and 10,000
{key_generator, {uniform_int, 10000}}.

% Use a randomly selected integer between 1 and 10,000, as binary.
{key_generator, {int_to_bin, {uniform_int, 10000}}}.

% Use a pareto distributed integer between 1 and 10,000; values < 2000
% will be returned 80% of the time.
{key_generator, {pareto_int, 10000}}.

value_generator

The generator function to use for creating values. Generators are defined in basho_bench_valgen.erl. Available generators include:

  • {fixed_bin, Size}) - generates a random binary ofSize` bytes. Every binary is the same size, but varies in content.
  • {exponential_bin, MinSize, Mean}) - generates a random binary which has an exponentially distributed size. Most values will be approximatelyMinSize+Mean` bytes in size, with a long tail of larger values.
  • {uniform_bin, MinSize, MaxSize}) - generates a random binary which has an evenly distributed size betweenMinSizeand MaxSize`.
  • {function, Module, Function, Args}) - specifies an external function that should return a value generator function. The worker Idwill be prepended toArgs` when the function is called.

The default value generator is {value_generator, {fixed_bin, 100}}.

Examples:

% Generate a fixed size random binary of 512 bytes
{value_generator, {fixed_bin, 512}}.

% Generate a random binary whose size is exponentially distributed
% starting at 1000 bytes and a mean of 2000 bytes
{value_generator, {exponential_bin, 1000, 2000}}.

rng_seed

The initial random seed to use. This is explicitly seeded, rather than seeded from the current time, so that a test can be run in a predictable, repeatable fashion.

Default is {rng_seed, {42, 23, 12}}.

% Seed to {12, 34, 56}
{rng_seed, {12, 34, 56}.

log_level

The log_level setting determines which messages Basho Bench will log to the console and to disk.

The default level is debug.

Valid levels
debug
info
warning
error

report_interval

How often, in seconds, the stats process should write histogram data to disk. The default is 10 seconds.

test_dir

The directory in which result data is written. The default is /tests.

basho_bench_driver_riakclient Settings

These configuration settings apply to the basho_bench_driver_riakclient driver.

riakclient_nodes

List of Riak nodes to use for testing.

{riakclient_nodes, ['riak1@127.0.0.1', 'riak2@127.0.0.1']}.

The Erlang cookie to use to connect to Riak clients. The default is riak.

{riakclient_cookie, riak}.

riakclient_mynode

The name of the local node. This is passed into net_kernel:start/1.

{riakclient_mynode, ['basho_bench@127.0.0.1', longnames]}.

riakclient_replies

This value is used for R-values during a get operation, and W-values during a put operation.

% Expect 1 reply.
{riakclient_replies, 1}.

riakclient_bucket

The Riak bucket to use for reading and writing values. The Default is <<"test">>.

% Use the "bench" bucket.
{riakclient_bucket, <<"bench">>}.

basho_bench_driver_riakc_pb Settings

riakc_pb_ips

A list of IP addresses to connect the workers to. A random IP will be chosen for each worker.

The default is {riakc_pb_ips, [{127,0,0,1}]}

% Connect to a cluster of 3 machines
{riakc_pb_ips, [{10,0,0,1},{10,0,0,2},{10,0,0,3}]}

riakc_pb_port

The port on which to connect to the PBC interface.

The default is {riakc_pb_port, 8087}

riakc_pb_bucket

The bucket to use for testing.

The default is {riakc_pb_bucket, <<"test">>}

basho_bench_driver_http_raw Settings

http_raw_ips

A list of IP addresses to connect the workers to. Each worker makes requests to each IP in a round-robin fashion.

The default is {http_raw_ips, ["127.0.0.1"]}

% Connect to a cluster of machines in the 10.x network
{http_raw_ips, ["10.0.0.1", "10.0.0.2", "10.0.0.3"]}.

http_raw_port

Select the default port to connect to for the HTTP server.

The default is {http_raw_port, 8098}.

% Connect on port 8090
{http_raw_port, 8090}.

http_raw_path

The base path to use for accessing Riak, usually "/riak/<bucket>".

The default is {http_raw_path, "/riak/test"}.

% Place test data in another_bucket
{http_raw_path, "/riak/another_bucket"}.

http_raw_params

Additional parameters to add to the end of the URL. This can be used to set the r/w/dw/rw parameters as desired.

The default is {http_raw_params, ""}.

% Set R=1, W=1 for testing a system with n_val set to 1
{http_raw_params, "?r=1&w=1"}.

http_raw_disconnect_frequency

How often, in seconds or number of operations, the HTTP clients (workers) should forcibly disconnect from the server.

The default is {http_raw_disconnect_frequency, infinity} (which means that Basho Bench should never forcibly disconnect).

% Disconnect after 60 seconds
{http_raw_disconnect_frequency, 60}.

% Disconnect after 200 operations
{http_raw_disconnect_frequency, {ops, 200}}.

Custom Driver

A custom driver must expose the following callbacks.

% Create the worker
% ID is an integer
new(ID) -> {ok, State} or {error, Reason}.

% Run an operation
run(Op, KeyGen, ValueGen, State) -> {ok, NewState} or {error, Reason, NewState}.

See the existing drivers for more details.