Skip to content

configuration.md

Latest commit

 

History

History
123 lines (100 loc) · 18.7 KB

configuration.md

File metadata and controls

123 lines (100 loc) · 18.7 KB

Loader configuration file format

Parameter name Data type Possible values Default value Description
Seed int64 any 42 Seed for specification generator (for reproducibility)
Platform string Knative, OpenWhisk, AWSLambda, Dirigent Knative The serverless platform the functions will be executed on
DirigentConfigPath 1 string N/A "" Path to the Dirigent configuration file
InvokeProtocol string grpc, http1, http2 N/A Protocol to use to communicate with the sandbox
YAMLSelector string wimpy, container, firecracker container Service YAML depending on sandbox type
EndpointPort int > 0 80 Port to be appended to the service URL
RpsTarget int >= 0 0 Number of requests per second to issue
RpsColdStartRatioPercentage int >= 0 && <= 100 0 Percentage of cold starts out of specified RPS
RpsCooldownSeconds 2 int > 0 0 The time it takes for the autoscaler to downscale function (higher for higher RPS)
RpsRuntimeMs int >= 0 0 Requested execution time
RpsMemoryMB int >= 0 0 Requested memory
RpsIterationMultiplier int >= 0 0 Iteration multiplier for RPS mode
TracePath 3 string string data/traces/example Folder with Azure trace dimensions (invocations.csv, durations.csv, memory.csv) or "RPS"
Granularity string minute, second minute Granularity for trace interpretation4
OutputPathPrefix string any data/out/experiment Results file(s) output path prefix
IATDistribution string exponential, exponential_shift, uniform, uniform_shift, equidistant exponential IAT distribution5
CPULimit string 1vCPU, GCP 1vCPU Imposed CPU limits on worker containers (only applicable for 'Knative' platform)6
ExperimentDuration int > 0 1 Experiment duration in minutes of trace to execute excluding warmup
WarmupDuration int > 0 0 Warmup duration in minutes(disabled if zero)
IsPartiallyPanic bool true/false false Pseudo-panic-mode only in Knative
EnableZipkinTracing bool true/false false Show loader span in Zipkin traces
EnableMetricsScrapping bool true/false false Scrap cluster-wide metrics
MetricScrapingPeriodSeconds int > 0 15 Period of Prometheus metrics scrapping
GRPCConnectionTimeoutSeconds int > 0 60 Timeout for establishing a gRPC connection
GRPCFunctionTimeoutSeconds int > 0 90 Maximum time given to function to execute7
DAGMode bool true/false false Generates DAG workflows iteratively with functions in TracePath 8. Frequency and IAT of the DAG follows their respective entry function, while Duration and Memory of each function will follow their respective values in TracePath.
EnableDAGDataset bool true/false true Generate width and depth from dag_structure.csv in TracePath9
Width int > 0 2 Default width of DAG
Depth int > 0 2 Default depth of DAG
VSwarm bool true/false false Execute vSwarm functions from mapper_output.json

InVitro can cause failure on cluster manager components. To do so, please configure the cmd/failure.json. Make sure that the node on which you run InVitro has SSH access to the target node.

Parameter name Description
FailureEnabled Toggle to enable this feature
FailAt Time in seconds since the beginning of the experiment when to trigger a failure
FailComponent Which component to fail (choose from 'control_plane', 'data_plane', 'worker_node')
FailNode Which node(s) to fail (specify separated by blank space)

Dirigent configuration

Parameter name Data type Possible values Default value Description
Backend string containerd, firecracker, dandelion containerd The backend used in Dirigent
DirigentControlPlaneIP string N/A N/A IP address of the Dirigent control plane (for function deployment)
BusyLoopOnSandboxStartup bool true/false false Enable artificial delay on sandbox startup
PrepullMode string all_sync, all_async, one_sync, one_async, none none Prepull image before starting experiments sync or async
AsyncMode bool true/false false Enable asynchronous invocations in Dirigent
AsyncResponseURL string N/A N/A URL from which to collect invocation responses
AsyncWaitToCollectMin int >= 0 0 Time after experiment ends after which to collect invocation results
RpsImage string N/A N/A Function image to use for RPS experiments
RpsRequestedGpu int >= 0 0 Number of gpus requested from Dirigent
RpsFile 3 string N/A N/A If given the payload is read from this file
RpsDataSizeMB 3 float64 >= 0 0 If no rps file is given this amount of random data is generated (same for all requests)
Workflow 4 bool true/false false Send workflow requests to Dirigent
WorkflowConfigPath 5 string N/A N/A Path to the configuration file for the workflow requests (see below)

3 Currently used only when requesting gpus (RpsRequestedGpu > 0) and ignored otherwise.

4 Only supported for backend dandelion.

5 Required only when Workflow is set to true.

Workflow configuration

Parameter name Data type Description
Name string Name to be used in the registration request.
Functions []WorkflowFunction Functions used in the composition(s).
Compositions []CompositionConfig Compositions defined in the workflow description.

WorkflowFunction

Parameter name Data type Description
FunctionName string Function name used in the workflow.
FunctionPath string Path to the binary located on the worker.
NumArgs int Number of input sets.
NumRets int Number of output sets.

CompositionConfig

Parameter name Data type Description
Name string Composition name.
InData 3 [][]string First dimension are the input sets, second one are the items (per set).

3 Prepend %path= to load the content from a local file path. Used empty string to use an empty input item.

Footnotes

  1. Required only when the Platform is Dirigent. ↩

  2. It is recommended that the first 10% of cold starts are discarded from the experiment results for low cold start RPS. ↩

  3. To run RPS experiments replace the path with RPS. ↩ ↩2 ↩3 ↩4 ↩5 ↩6

  4. The second granularity feature interprets each column of the trace as a second, rather than as a minute, and generates IAT for each second. This feature is useful for fine-grained and precise invocation scheduling in experiments involving stable low load. ↩ ↩2 ↩3

  5. _shift modifies the IAT generation in the following way: by default, generation will create first invocation in the beginning of the minute, with _shift modifier, it will be shifted inside the minute to remove the burst of invocations from all the functions. ↩ ↩2 ↩3

  6. Limits are set by resource->limits->CPU in the service YAML. 1vCPU means limit of 1CPU is set, at the same time execution is also limited by the container concurrency limit of 1. GCP means limits are set to multiples of 1/12th of vCPU, based on the memory consumption of the function according to this table for Google Cloud Functions. ↩

  7. Function can execute for at most 15 minutes as in AWS Lambda; https://aws.amazon.com/about-aws/whats-new/2018/10/aws-lambda-supports-functions-that-can-run-up-to-15-minutes/ ↩

  8. The generated DAGs consist of unique functions. The shape of each DAG is determined either Width,Depth or calculated based on EnableDAGDAtaset. ↩

  9. A data sample of DAG structures has been created based on past Microsoft Azure traces. Width and Depth are determined based on probabilities of this sample. ↩