tracebench

README

tracebench

Run deterministic workloads against a Wasimoff broker. It can either use a function generator given configured distributions or follow trace from a Huawei dataset.

go build
./tracebench -help

• Without -broker ... the tool runs in a dry-run mode, to check if your workload is as desired, without actually sending any tasks.
• Use -wait to wait for all in-flight tasks at the end of the configured duration. Otherwise the measurement is stopped abruptly and remaining tasks are cancelled.

The tool writes a file with event traces for each task in a varint-length encoded protobuf file (like tracebench-1765548497-309160.pb). You can either compile the Protobuf message definitions for your language of choice or convert it into a JSONL (each task is one JSON object per line):

go run ./proto2jsonl < tracebench.pb

funcgen

Runs a YAML file with configured workloads. See example_funcgen.yaml for possible options.

./tracebench -funcgen workloads.yaml -wait -broker http://localhost:4080/

tracer

This mode uses the FaaS traces from Huawei's 2023 data release from their private cloud. Use the downloader in dataset/download.py to retrieve the necessary data files and configure the workloads as shown in example_csvtrace.yaml. You may use dataset/viewer.py or the original authors' Notebooks to find suitable function columns for your use-case.

./tracebench -trace huawei.yaml -wait -broker http://localhost:4080/