App-Layer Throughput: Your Software Can Be the Bottleneck

Summary

Even with a tuned OS and 100 Gbps port, application architecture may limit throughput. Single-threaded tools or conservative defaults (e.g., worker counts, buffer sizes, TLS/SSH overhead) often cap performance.

Common examples & fixes

• File transfer: SCP/SFTP are single-threaded. Use bbcp or rsync with parallel streams.
• Web services / APIs: Increase worker processes/threads, buffer sizes, keep-alive.
• Databases: Enable parallel query execution, scale out consumers.
• Virtualization: Ensure virtio/vmxnet3 drivers and offloads are configured.

Validation workflow

1. Prove host ↔ host raw throughput with iPerf3.
2. Add the application and watch where performance drops.
3. Profile CPU, latency, and per-flow rates; scale parallelism.
4. For commercial/open source software packages, review documentation for performance considerations carefully.

References

• bbcp overview: parallel streams over SSH for high-speed transfer. ESnet/US Dept. of Energy
• ESnet: Linux Host Tuning general guide under Host Tuning → Linux. oai_citation:16‡fasterdata.es.net
• ESnet: 100G Benchmarking & Other Tuning pages. oai_citation:17‡fasterdata.es.net