- Java 9 or later
- sbt
- make
Execute sbt assembly or just plain make
Copy the resulting executable target/scala-2.12/perftest.jar to some
machine close to the server(s) to test, then execute something like
this, perhaps in multiple copies in parallel, as the users parameter
does not add load linearly for higher numbers. Replace the
test.EXAMPLE.COM and token string with relevant name and token for your
server to be tested. You may need to run the test client on multiple
machines in order to generate more traffic than one network link can
handle, if that is part of the test.
java -Dbulksize=1000 -Ddatasources=50 -Dbaseurls=https://test.EXAMPLE.COM -Dusers=100 -Dtoken=<SOME-INGEST-TOKEN> -jar ./perftest.jar -s com.humio.perftest.HECSimulation
java -Dbulksize=1000 -Ddatasources=50 -Dbaseurls=https://test.EXAMPLE.COM -Dusers=100 -Dtoken=<SOME-INGEST-TOKEN> -Dtemplate=templates/test.ssp -jar ./perftest.jar -s com.humio.perftest.HECSimulation
java -Dbulksize=1000 -Ddatasources=50 -Dbaseurls=https://test.EXAMPLE.COM -Dusers=100 -Drandomness=3 -Dtoken=<SOME-INGEST-TOKEN> -jar ./perftest.jar -s com.humio.perftest.HECRandomnessSimulation
java -Dbulksize=1000 -Ddatasources=50 -Dbaseurls=https://test.EXAMPLE.COM -Dusers=100 -Dtoken=<SOME-INGEST-TOKEN> -jar ./perftest.jar -s com.humio.perftest.FilebeatSimulation
For HECSimulation the URL should be $BASEURL/api/v1/ingest/hec.
For the FilebeatSimulation and FixedRateIngestSimulation the URL should be $BASEURL/api/v1/ingest/elastic-bulk.
$BASEURL should be the URL containing the Humio host. For example https://cloud.humio.com.
docker run \
-e PERF_TIME=300 \
-e PERF_TENS_GB_PER_DAY=10 \
-e HUMIO_TOKEN=<SOME-INGEST-TOKEN> \
-e HUMIO_BASE_URL=<URL to Humio ingest endpoint> \
-e PERF_SIMULATION=FixedRateIngestSimulation \
humio/humio-ingest-load-test:latest
docker run \
-e PERF_TIME=300 \
-e PERF_USERS=1000 \
-e PERF_DATASOURCES=50 \
-e PERF_BULK_SIZE=1000 \
-e HUMIO_TOKEN=<SOME-INGEST-TOKEN> \
-e HUMIO_BASE_URL=<URL to Humio ingest endpoint> \
-e PERF_SIMULATION=<HECSimulation or FilebeatSimulation> \
humio/humio-ingest-load-test:latest
docker run \
-e PERF_TIME=300 \
-e PERF_USERS=1000 \
-e PERF_DATASOURCES=50 \
-e PERF_BULK_SIZE=1000 \
-e HUMIO_TOKEN=<SOME-INGEST-TOKEN> \
-e HUMIO_BASE_URL=<URL to Humio ingest endpoint> \
-e PERF_SIMULATION=HECRandomnessSimulation \
-e RANDOMNESS=3 \
humio/humio-ingest-load-test:latest
docker run \
-e PERF_TIME=300 \
-e PERF_USERS=1000 \
-e PERF_DATASOURCES=50 \
-e PERF_BULK_SIZE=1000 \
-e HUMIO_TOKEN=<SOME-INGEST-TOKEN> \
-e HUMIO_BASE_URL=<URL to Humio ingest endpoint> \
-e PERF_SIMULATION=HECTemplateSimulation \
-e TEMPLATE=templates/test.ssp \
humio/humio-ingest-load-test:latest
# If you want to use a custom templates and data you need to mount a local volume into the container.
docker run \
-e PERF_TIME=300 \
-e PERF_USERS=1000 \
-e PERF_DATASOURCES=50 \
-e PERF_BULK_SIZE=1000 \
-e HUMIO_TOKEN=<SOME-INGEST-TOKEN> \
-e HUMIO_BASE_URL=<URL to Humio ingest endpoint> \
-e PERF_SIMULATION=HECTemplateSimulation \
-e TEMPLATE=templates/my-template.ssp> \
-v /Users/me/humio-load-test-templates:/humio-ingest-load-test/templates/my-template.ssp> \
humio/humio-ingest-load-test:latest
docker run \
-e PERF_SEARCH_QUERY="count()" \
-e PERF_SEARCH_DURATION=24hours \
-e HUMIO_TOKEN=<SOME-INGEST-TOKEN> \
-e HUMIO_BASE_URL=<URL to Humio> \
-e PERF_SIMULATION=QuerySimulation \
humio/humio-ingest-load-test:latest
The goal of HECTemplateSimulation is to make dynamic content generation easier, as well as generate data that is realistically compressible and, ideally, quasi-meaningfully queryable for demonstration purposes.
Toward this end, HECTemplateSimulation provides the ability to use a specified .ssp file, interpreted by the Scalate engine; in this specific instance, the SSP implementation is used, which is similar to Velocity, JSP, Erb, etc.
Along with the ability to template, HECTemplateSimulation adds the ability to generate different types of data according to a specified distribution (examples below).
SSP (so-called "Scala Server Pages") directives are documented in full here.
Below I will walk you through the different sections, and mechanics, of template handling. See the file templates/test.ssp for a full working example.
Templates require a number of imports to function correctly. More may be required if you implement custom functionality beyond that supported by these typical libraries:
import com.humio.perftest._
import play.api.libs.json._
import org.fusesource.scalate._
import org.fusesource.scalate.RenderContext._
Templates typically use an init block to create and register data generators and one-time values. This block of the template is only run once upon initialization. It is implemented very simply, switched on a context variable init:
if (init) {
data.register("httpMethod",
new ArraySampler(
data.distributions.exponential,
Array("GET", "POST", "PUT", "DELETE", "OPTION")
)
)
// ....
}
All code outside the init block between <% and %> (see the SSP & Scalate reference docs for more information on this and other template directives) is executed on every data generation request. This is the section where the actual data generation from the template takes place. For example, with the templates/test.ssp example, this section is outside the init block and is executed on every generation call:
val sourceType = "applog"
val logTs = data.timestamp
val tag = "tag" + data.sample("tagNum")
val source = "file" + data.sample("sourceNum")
val viewId = data.sha1(data.sample("viewAndQueryId")).substring(0,10)
val queryId = "IQ-" + viewId.substring(0,5).reverse + viewId.substring(5).reverse
var logLine: String = capture {
%>
${logTs} requests 11 - route=humio method=<%= data.sample("httpMethod") %>, remote=10.0.<%= data.sample("host1") %>.<%= data.sample("host2") %>:<%= data.sample("port") %> uri=http://newcloud<%= data.sample("cloudHost") %>:8080/api/v1/internal/views/${viewId}/queryjob/${queryId}, status=<%= data.sample("status") %>, time=<%= data.sample("time") %>, internal=<%= data.sample("internal") %>, contentLength=<%= data.sample("contentLength") %>
<%
}.trim // capture
val jsonObj = Json.obj(
"source" -> source,
"sourcetype" -> sourceType,
"event" -> logLine,
"time" -> logTs,
"fields" -> Json.obj(
"host" -> "perftesthost",
"source2" -> source
)
)
val output = Json.toJson(jsonObj)
%>${output.toString.trim}
Simple data generation methods provide things like timestamps, etc. Sampled data generators must be created and registered. Their function is to select from sets of elements with specified probabilities.
There are a handful of functions available to generate simple types of data:
| Name | Implementation |
|---|---|
data.timestamp |
Current timestamp, via "%.6f".format(ZonedDateTime.now().toInstant.toEpochMilli.toDouble / 1000d) |
data.iso8601Timestamp |
Current timestamp, via SimpleDateFormat format yyyy-MM-dd'T'HH:mm'Z' |
data.sha1(String) |
SHA1 hash of given string. Useful for generating identifiers from other data types. |
For example, let's say you have a log line you would like to generate similar log lines of:
2021-01-01T0:0:0.0Z requests 11 - route=humio method=GET, remote=10.0.1.10, ...
We want to replace the values for the method and remote key-value pairs with data that is plausible. The first step is to think about how the log line works: "What is the distribution of HTTP request methods for this endpoint?". For this example, let's say it's predominantly GET, with POST a distant second, and the rest typically representative of error or malformed calls. For the purposes of this example, let's say it's best represented by an exponential distribution. Highly accurate distribution estimation is not generally necessary to achieve the aims of this project and its methods are beyond the scope of this document.
So, for method, to sample from an array of values with exponentially decreasing probability (more later on his this mapping is achieved), you can create an ArraySampler and register it with the name httpMethod:
data.register("httpMethod",
new ArraySampler(
data.distributions.exponential,
Array("GET", "POST", "PUT", "DELETE", "OPTION")
)
)
The above code creates an ArraySampler that maps exponentially decreasing probabilities to successive elements in the provided array.
Now, the remote element. To make this more interesting to query when combined with other generated elements, let's break the field into 10.0.host1.host2, where the distributions of host1 and host2, while different, combine to form a useful structure of IP addresses in the aggregate. For example:
data.register("host1", new IntSampler(data.distributions.exponential, 0, 10))
data.register("host2", new IntSampler(data.distributions.uniform, 0, 20))
The first generator, host1, is defined as a sampler over the integer range 0 to 10 with decreasing exponential probability (i.e., 0 is most likely, 10 least). The second generator, host2, is defined as a sampler over the integer range 0 to 20, with uniform probability.
The available samplers and distributions are described in complete below.
Once a sampler (e.g., ArraySampler) has been created and registered, it can be sampled from to produce a value. Following along with the templates/test.ssp example, the example log line we're trying to create a generator for becomes the following:
${logTs} requests 11 - route=humio method=<%= data.sample("httpMethod") %>, remote=10.0.<%= data.sample("host1") %>.<%= data.sample("host2") %>, ...
The relevant part in this context is this: <$= data.sample("httpMethod") %>. The <%= and %> tags say "execute the code within this block and output the value. The data.sample("httpMethod") code requests a sample from the sampler registered under the name httpMethod, which we created and registered above. The same concept applies to the host1 and host2 samplers.
Note: CSVSampler is a sampler that exposes a sampleRow function as well as a sample function. sampleRow samples a row of values from the provided CSV file that can then be access by index, e.g., someVar(1). Using sample with a CSVSampler simply returns the value in the first column of the sampled row.
| Name | Description | Signature |
|---|---|---|
| ArraySampler | Samples from an array of Strings. | new ArraySampler(<distribution>, Array("value", ...)) |
| IntSampler | Samples from an increasing range of integers. | new IntSampler(<distribution>, <lower>, <upper>) |
| RealSampler | Samples from an increasing range of doubles. | new RealSampler(<distribution>, <lower>, <upper>) |
| CSVSampler | Samples rows from a CSV file. | new CSVSampler(<distribution>, "<csv filename>") |
Distributions are implemented with the Apache Commons Math Library. Since they are distributions over the reals, the following method was used to map probabilities (see SimUtils for further information):
abstract class Sampleable(distribution: RealDistribution) {
val minProbability = 0.0001
val maxProbability = 0.999999
val icp0 = distribution.inverseCumulativeProbability(minProbability)
val icp1 = distribution.inverseCumulativeProbability(maxProbability)
val icpDif = icp1 - icp0
def sampleDistribution = {
val sample = distribution.sample()
if (sample < icp0) 0
else if (sample > icp1) 1
else (sample - icp0) / icpDif
}
def sample: String
}
With the mapping following the pattern similar to:
class ArraySampler(distribution: RealDistribution, values:Array[String]) extends Sampleable(distribution = distribution) {
override def sample: String = values(((values.length-1).toDouble * sampleDistribution).round.toInt)
}
The actual distribution sampling, distribution.sample(), is implemented by Inverse transform sampling.
Distribution specification has been simplified with default values, which are described below. If you require a specific configuration, see TemplateHelper.distributions in SimUtils.
| Name | Description |
|---|---|
| data.distributions.exponential | Exponential distribution, default mean=1 |
| data.distributions.normal | Normal distribution, default mean=0, scale=1 |
| data.distributions.uniform | Uniform distribution |
| data.distributions.logNormal | Log-normal distribution, default scale=0, shape=1 |
In a pinch, you can import import org.apache.commons.math3.distribution._ in your template and construct a distribution directly if you don't want to add the specific distribution and associated default parameters to the default options, e.g.,
data.register("contentLength",
new IntSampler(
new BetaDistribution(0.5, 0.5),
0, 256000))
All available distributions are subclasses of AbstractRealDistribution.
All distribution sampling uses the Well19937c generator.
Let's say you have a pair of fields being generated that must be correlated, e.g., a pair of fields named rcode_name and rcode, where they both hold a value that relates to the other. A good way to do this is to use a CSVSampler that references a CSV file that maps one to the other, e.g.,
NOERROR,0
SERVFAIL,2
NXDOMAIN,3
REFUSED,5
With this input to the CSVSampler, along with an appropriate distribution, you can use sampleRow, e.g.,
// init:
data.register("rcode", new CSVSampler(data.distributions.exponential, "templates/data/dns_rcode.samples.csv"))
// sample:
val rcodeRow = data.sampleRow("rcode")
val rcodeName = rcodeRow(0)
val rcode = rcodeRow(1)
For a real-world example of this, see the corelight-dns.ssp template.
Let's say you have a variable that you're unsure how to model, but you do have many observations of that variable (the "sample frame"). One method of sampling this variable is to load a CSV file with all the observations -- do not process this list of values in any way! -- into a CSVSampler with a uniform distribution, e.g.,
// init:
data.register("TTL", new CSVSampler(data.distributions.uniform, "templates/data/dns_ttls0.csv"))
// sample:
val ttl = data.sample("TTL")
The bigger the sample frame, the more the sampled data will be representative of the underlying distribution.
Notes:
- All
CSVSamplerdata is stored in memory: be mindful and plan accordingly when using large sets of observations. - While there are many opportunities for significant error when using this method, for simple data generation tasks it can still be useful.
There is a command-line tool to test templates. It will generate data corresponding to 10 executions of the template. Run sbt, then runMain com.humio.perftest.TemplateTest <filename>. For example:
sbt:humio-ingest-load-test> runMain com.humio.perftest.TemplateTest templates/test.ssp
(...)
{"source":"file8","sourcetype":"applog","event":"1613578255.420000 requests 11 - route=humio method=GET, remote=10.0.0.12:33392 uri=http://newcloud8:8080/api/v1/internal/views/tlifxqsNyC/queryjob/IQ-xfiltCyNsq, status=200, time=0.07, internal=true, contentLength=3308","time":"1613578255.420000","fields":{"host":"perftesthost","source2":"file8"}}
{"source":"file11","sourcetype":"applog","event":"1613578255.574000 requests 11 - route=humio method=GET, remote=10.0.2.6:37436 uri=http://newcloud6:8080/api/v1/internal/views/tlifxqsNyC/queryjob/IQ-xfiltCyNsq, status=200, time=0.01, internal=true, contentLength=1336","time":"1613578255.574000","fields":{"host":"perftesthost","source2":"file11"}}
{"source":"file8","sourcetype":"applog","event":"1613578255.575000 requests 11 - route=humio method=GET, remote=10.0.0.6:43388 uri=http://newcloud7:8080/api/v1/internal/views/tlifxqsNyC/queryjob/IQ-xfiltCyNsq, status=200, time=0.17, internal=true, contentLength=5218","time":"1613578255.575000","fields":{"host":"perftesthost","source2":"file8"}}
{"source":"file7","sourcetype":"applog","event":"1613578255.575000 requests 11 - route=humio method=GET, remote=10.0.0.24:35596 uri=http://newcloud4:8080/api/v1/internal/views/tlifxqsNyC/queryjob/IQ-xfiltCyNsq, status=200, time=0.11, internal=true, contentLength=7093","time":"1613578255.575000","fields":{"host":"perftesthost","source2":"file7"}}
{"source":"file6","sourcetype":"applog","event":"1613578255.576000 requests 11 - route=humio method=POST, remote=10.0.0.5:24921 uri=http://newcloud1:8080/api/v1/internal/views/tlifxqsNyC/queryjob/IQ-xfiltCyNsq, status=200, time=0.02, internal=true, contentLength=3969","time":"1613578255.576000","fields":{"host":"perftesthost","source2":"file6"}}
{"source":"file11","sourcetype":"applog","event":"1613578255.576000 requests 11 - route=humio method=GET, remote=10.0.1.28:30394 uri=http://newcloud8:8080/api/v1/internal/views/NWoZK3kTsE/queryjob/IQ-KZoWNEsTk3, status=200, time=0.04, internal=true, contentLength=535","time":"1613578255.576000","fields":{"host":"perftesthost","source2":"file11"}}
{"source":"file7","sourcetype":"applog","event":"1613578255.577000 requests 11 - route=humio method=GET, remote=10.0.1.26:31617 uri=http://newcloud10:8080/api/v1/internal/views/tlifxqsNyC/queryjob/IQ-xfiltCyNsq, status=200, time=0.04, internal=true, contentLength=1699","time":"1613578255.577000","fields":{"host":"perftesthost","source2":"file7"}}
{"source":"file9","sourcetype":"applog","event":"1613578255.577000 requests 11 - route=humio method=POST, remote=10.0.0.44:56138 uri=http://newcloud5:8080/api/v1/internal/views/tlifxqsNyC/queryjob/IQ-xfiltCyNsq, status=200, time=0.13, internal=true, contentLength=3224","time":"1613578255.577000","fields":{"host":"perftesthost","source2":"file9"}}
{"source":"file7","sourcetype":"applog","event":"1613578255.578000 requests 11 - route=humio method=GET, remote=10.0.2.17:35097 uri=http://newcloud10:8080/api/v1/internal/views/tlifxqsNyC/queryjob/IQ-xfiltCyNsq, status=200, time=0.05, internal=true, contentLength=6346","time":"1613578255.578000","fields":{"host":"perftesthost","source2":"file7"}}
{"source":"file9","sourcetype":"applog","event":"1613578255.578000 requests 11 - route=humio method=GET, remote=10.0.0.12:56221 uri=http://newcloud11:8080/api/v1/internal/views/tlifxqsNyC/queryjob/IQ-xfiltCyNsq, status=200, time=0.11, internal=true, contentLength=5762","time":"1613578255.578000","fields":{"host":"perftesthost","source2":"file9"}}
{"source":"file4","sourcetype":"applog","event":"1613578255.579000 requests 11 - route=humio method=GET, remote=10.0.0.13:36705 uri=http://newcloud13:8080/api/v1/internal/views/tlifxqsNyC/queryjob/IQ-xfiltCyNsq, status=204, time=0.0, internal=true, contentLength=709","time":"1613578255.579000","fields":{"host":"perftesthost","source2":"file4"}}