Skip to content

Latest commit

 

History

History
147 lines (109 loc) · 7.42 KB

README.md

File metadata and controls

147 lines (109 loc) · 7.42 KB
Raw

Configure Gadget

A Gadget user configures Gadget by setting three main elements:

  • Event source - There can be two event sources
  • operators type
  • KV store

The events source tells Gadget the sources of events. In Gadget, events can come from a trace (e.g., Google borg traces), or Gadget itself can generate events.

The operator types (e.g., tumbling.keyed.incremental) determine the operator for which Gadget generates state accesses load.

The KV store determines the KV store that Gadget evaluates using the generated state accesses load. If the user chooses a file, the generated state accesses load will be dumped on the file named gadget.log

Each main component has several parameters that need to be configured by the user. The Gadget compiler helps users set the required parameters by outputting error messages.

Table of Contents

Event Source

Gadget needs one event source for single-input operators (e.g., tumbling and sliding windows) and two event sources for two-inputs operators (eg., sliding and tumbling joins) Gadget offers two types of event generators: 1) gadget 2) tracefile

When the event generator is gadget, it generates desired events. When the event generator is a tracefile, the events are read from a trace file.

Lets first set the event generator type (example: please see line 4 in file config.txt):

first.event.generator.type = gadget

Next, lets set the needed parameters based on the type of the event generator:

when event.generator.type is gadget

If the event generator type is set to gadget, we must configure the key popularity and event time distribution of events. We also need to configure the watermark frequency, the percentage of out-of-order events and the lateness threshold. Please see lines 6-18 of config.txt as an example.

The following lines set the key popularity of events. The key popularity can be set to constant, hotspot, zipf, sequential, and uniform.

#options(constant, hotspot, zipf, sequential, uniform)
first.event.generator.key.popularity.distrib.type = zipf
first.event.generator.key.popularity.distrib.params.s = 1
first.event.generator.key.popularity.distrib.params.keyspacesize= 10000000

The following lines set the distribution of event times.

#options(exponential, normal, constant, uniform)
first.event.generator.event.occurrence.gap.distrib.type = exponential
first.event.generator.event.occurrence.gap.distrib.params.lambda= 10

The following lines configure watermark frequency, the percentage of out-of-order events, and the lateness threshold

first.event.generator.watermark.frequency = 1
first.event.generator.outoforder.percentage = 2
first.event.generator.lateness.threshold = 3;

when event.generator.type is tracefile

If the event generator type is set to tracefile, the user needs to set the column indexes for event key, event time, and event type (indexes start from 0). For example, config2.txt (lines 4-7) sets an event generator of type tracefile. It uses task-input.txt as the event source (please find this trace under traces)

#options: (tracefile, gadget)
first.event.generator.type = tracefile
first.event.generator.key.index = 0
first.event.generator.event.time.index = 2
first.event.generator.event.type.index = 4
first.event.generator.file.path = tasks-input.txt
first.event.generator.watermark.frequency = 1

Set the operator

To configure the operators, users need to set the operator type, its required parameters, kay popularity, key size, and value size distribution of generated state access operations.

operator type

Gadget offers the following operators for which it generates state accesses workload: tumbling.all.incremental, tumbling.all.holistic, tumbling.keyed.incremental, tumbling.keyed.holistic, sliding.all.incremental, sliding.all.holistic, sliding.keyed.incremental, sliding.keyed.holistic, join.tumbling, join.sliding, join.interval, session.holistic ,session.incremental , continuous.join, flink.replayer, ycsb.replayer. Tumbling window operators need window.length and sliding window operators need both window.length and sliding.length to be set. For example, in config.txt (lines 27-30), the operator type us set to tumbling.keyed.incremental and window.length is set to 5 units of time.

operator.type= tumbling.keyed.incremental
expected.num.operations = 100000
window.length = 5
#sliding.length = 1

expected.num.operations determines the number of state access operations that will be generated.

key popularity

Gadget users can define the key popularity of state objects (ie., window IDs). In file config.txt key popularity is set to sequential, meaning that the key of the newly created window (state) is set to LastWindowID+1. Gadget Users do not need to change this unless exploring novel window management strategies.

#options(constant, hotspot, zipf, sequential, uniform) - no need to change this
operator.key.popularity.distrib.type = sequential
operator.key.popularity.distrib.params.keyspacesize = 1000000000

key and value sizes

Gadget users can set the key and value size of state accesses. For example, in file config.txt, the key and value size of state operations is constant and equal to 10 bytes.

#********  key size distribution (keys used to access state store)
#options(exponential, constant, uniform, normal)
operator.key.size.distrib.type = constant
operator.key.size.distrib.params.constant = 10

#******** value size distribution (values that are inserted to state store)
#options(exponential, constant, uniform, normal)
operator.value.size.distrib.type = constant
operator.value.size.distrib.params.constant = 10

Evaluate a KV store or dump on file

Finally, Gadget users can choose to dump the generated state workload on a file or evaluate a KV store performance using the generated state workload.

# options(file - rocksdb- berkeley - faster)
#If file is chosen, Gadget dumps the generated stat access operations on the file (wrapper.path/gadget.log)
#if you choose to evaluate a KV store, you need to install the KV store first. 
wrapper.type = file
wrapper.path = .

The distributions

Please note if you set the distribution of value or key sizes, popularity or event times, you may need to configure some extra parameters depending on the type of distribution. For example:

if you set the distribution to exponential, the parameter lambda needs to be defined.

If you set the distribution to constant, the parameter constant needs to be defined.

If you set the distribution to normal, the parameters sd and mean need to be defined.

If you set the distribution to uniform, the parameters a and b need to be defined, where a is the lower bound, and b is the upper bound.

If you set the distribution to hotspot, the parameters hotspotfraction and hotoperationfraction need to be defined.

If you set the distribution to zipf, the parameter s needs to be defined.