S3 Benchmark

Your Amazon S3 performance depends on 3 things:

1. Your distance to the S3 endpoint.
2. The size of your objects.
3. The number of parallel transfers you can make.

With this tool you can measure S3's performance from any location, using different thread counts and object sizes. Here's an example from a c5.4xlarge EC2 instance:

Usage

Download

macOS

curl -OL https://github.com/dvassallo/s3-benchmark/raw/master/build/darwin-amd64/s3-benchmark

Linux 64-bit x86

curl -OL https://github.com/dvassallo/s3-benchmark/raw/master/build/linux-amd64/s3-benchmark

Linux 64-bit ARM

curl -OL https://github.com/dvassallo/s3-benchmark/raw/master/build/linux-arm64/s3-benchmark

Credentials

This tool needs AWS credentials with full S3 permissions. If you run this on EC2, it will automatically use your EC2 instance profile. Otherwise it will try to find the credentials from the usual places.

Run

Make the file executable:

chmod +x s3-benchmark

Run a quick test (takes a few minutes):

./s3-benchmark

Or run the full test (takes a few hours):

./s3-benchmark -full

See this for all the other options.

Build

1. Install Go

   sudo apt-get install golang-go
   

   or

   sudo yum install go
   

   may work too.

2. Setup Go environment variables (Usually GOPATH and GOBIN) and test Go installation

3. Clone the repo

4. Install dep

   go get -u github.com/golang/dep/cmd/dep
   

5. Go to source directory and run dep ensure

6. Run go run main.go

S3 to EC2 Bandwidth

I ran this benchmark on all current generation EC2 instance types as of 2019-04-02. I put the data in an interactive spreadsheet that you can download here. Feel free to share it with your friends and colleagues.

Analysis

Here's the maximum download throughput I got from all 155 instance types:

Here's the performance of all instances with 32 MB objects (the legend is truncated, but all instances are plotted):

And here's the same chart with just the 3 outlier instances that have 50 Gigabit or more network bandwidth:

Here's a typical throughput profile showing how object size affects performance:

S3's 90th percentile time to first byte is typically around 20 ms regardless of the object size. However, small instances start to see elevated latencies with increased parallelization due to their limited resources. Here's the p90 first byte latency on a small instance:

And here's the p90 first byte latency on a larger instance:

Unlike the first byte latency, the time to last byte obviously follows the object size. S3 seems to deliver downloads at a rate of about 93 MB/s per thread, so this latency is a function of that and the first byte latency — at least until the network bandwidth gets saturated. Here's one example:

If you want to analyze the data further, you can download the spreadsheet with all the raw data and the interactive features shown in the screenshots. Here's a 45 second demo of what you can do with it.

License

This project is released under the MIT License.