UdS FS infra benchmarks

The main goal is to figure out what network latency is acceptable for NFS. Testing with netem might be a way to do so.

Workload benchmarks are:

• create N files
• read N files
• ls folder
• del N files
• sequential reads
• random read/writes
• sequential writes

Usage

Run automated benchmark that automatically sets the netem latency and measures all of the above:

python3 latency_benchmark.py benchmark <iface>
# optional with custom latency list:
python3 latency_benchmark.py benchmark <iface> "[0, 1000, 2000, 4000, 8000, 16000, 32000]"

The results are dumped to a <node_name>.csv file, and three plots (IOPS, bandwidth, and file operations) are stored in the working directory.

Exemplary, we can see from the bandwidth plot below, (which has a log y-axis) that the bandwidth is scaling exponentially with the latency:

Same with the IOPS:

The fileops on the other hand scale linearly (note, this is the only plot not in log scale!):