Skip to main content Link Search Menu Expand Document (external link)

Experiment data

This page describes how to obtain the raw data produced during an experiment. This includes corpora, crashes and logs from fuzzers. It does not include the metrics used to generate a report, that data is linked to in the report itself.

This page is written for users of the FuzzBench service. The same concepts apply to users running FuzzBench on their own. However, the Google Cloud Storage buckets will be different.

Getting the data

We store experiment data in the Google cloud storage bucket: “gs://fuzzbench-data”. Provides a web interface for browsing and downloading the experiment data.

We use gsutil for obtaining this data. Follow these instructions to install gsutil.

With gsutil you can do gsutil ls gs://fuzzbench-data/$EXPERIMENT_NAME/ to list files in cloud storage directories for your project and you can use gsutil cp to copy them.

Data layout

Let’s see what the layout of this data directory looks like:

│       oss-fuzz-$OSS_FUZZ_PROJECT-fuzzer-$FUZZER-hash-$DOCKER_HASH.txt  # Logs for an OSS-Fuzz benchmark build.
│       benchmark-$BENCHMARK-fuzzer-$FUZZER-hash-$DOCKER_HASH.txt  # Logs for a standard benchmark build.
│       ...
│       coverage-build-$BENCHMARK.tar.gz
│       ...
│   │
│   │   │
│   │   └───trial-$TRIAL_NUM
│   │   │   │
│   │   │   └───corpus
│   │   │   │       corpus-archive-0001.tar.gz
│   │   │   │       ...
│   │   │   │
│   │   │   └───crashes
│   │   │   │       crashes-0001.tar.gz
│   │   │   │       ...
│   │   │   │
│   │   │   └─results
│   │   │          fuzzer-log.txt
│   │   │          unchanged-cycles
│   │   │
│   │   └─...
│   │
│   └─ ... 
└───input # Contains source code and config files used to run experiment.

Corpus archive and unchanged-cycles

A trial is a run of a specific fuzzer on a specific benchmark. For each trial, FuzzBench archives the output_corpus used in every 15 minutes, this is sometimes referred to as a “snapshot” (or “cycle”). As an optimization, if the directory hasn’t changed since the last cycle, the archiving step is skipped and the cycle number is added to the unchanged-cycles file. For some fuzzers like AFL, we ignore frequently changing non-corpus files like fuzzer_stats when determining if the corpus changed since last cycle. Because fuzzer_stats is included in the archive, you can obtain stats for an AFL-based fuzzer using the last archive in a trial.


Though FuzzBench doesn’t use crashes for measuring performance, it does save them. When FuzzBench measures the coverage of a corpus snapshot, if it encounters any crashes it adds them to the crashes archive for that cycle.


The stdout and stderr from running a fuzzer.