The purpose of this experiment is to evaluate the performance of different storage solutions. In particular, we are interested in meausuring the number of IOPS and the bandwidth and looking at the tradeoff between the two.

To run the experiment on a local disk no real setup is required. A file will be created at /tmp/file.dat (change this if /tmp is a ramdisk) and the tests will run against this file.

To run the experiment on S3 you will need to create a bucket and the appropriate credentials. We construct the S3 object store using AmazonS3Builder::from_env() and so you will need to ensure the credentials are accessible in the environment.

Specifically, we used IAM profiles assigned to an EC2 instance for authorization. S3 results and S3 express results were generated on a c7gn.8xlarge instance.

To run the experiment, simply run one of the following commands:

# Tests on a local disk
cargo run --release --bin local
# Tests on s3
cargo run --release --bin s3

Results are written to a file named disk_results.csv or s3_results.csv depending on the storage solution used.

To generate a chart you can run the following command:

# Must run this in the root of the repository
python3 scripts/plot.py

This will generate a chart named chart.png in the root of the repository.

We have checked in the results of the experiment that were used in the paper. Here are some notes on how these results were generated.

• The NVMe results were generated on a local desktop with a Samsung 970 EVO Plus 2TB drive.
• The S3 results were generated on a c7gn.8xlarge instance
• The S3 express results were generated on a c7gn.8xlarge instance
• The S3 express results required modifying the s3 script to add the with_s3_express(true) call and changing the bucket name. We renamed the results file to s3_express_results.csv to avoid overwriting the original results.
• The HDD results were generated on a local desktop with a an ancient WDC WD10EZEX-08W.
• The HDD results were renamed to hdd_results.csv to avoid overwriting the original results.