Homa-Class Latency — Keep TCP For Clients and Servers

WarpRATE™’s PFQD-DC™ adds completion-aware scheduling to PFQD™ and TCP. Thereby delivering short-message latency competitive with Homa while preserving existing application networking interfaces.

In extensive simulations PFQD-DC ranged between 50-120% Homa performance. Deployment is via scheduling and packet marking on hosts and top-of-rack switches thereby allowing applications to keep TCP’s familiar interfaces and flow semantics.

Request a Demo Explore WarpEngine

Challenge

  • Distributed services depend on timely communication. When an operation waits for multiple responses, a delayed message can hold up the entire operation.
  • Short messages and bulk transfers often compete for the same host and switch queues. Adding bandwidth does not directly address which queued traffic gets served next.
  • Homa demonstrates the benefits of scheduling messages for rapid completion. Adopting it also introduces a different transport and messaging interface — creating non-trivial integration work for existing datacenter applications.

Approach

  • PFQD-DC extends Badu Networks’ per-flow scheduler with a Completion Tier. Sending hosts mark packets according to how much data remains queued behind them, allowing the scheduler to favor flows with less queued work remaining.
  • Scheduling runs at both the sending host and the top-of-rack switch. Completion priority has a bounded share of service, keeping other flows moving while preserving packet order within each flow.
  • The evaluated configuration retains TCP and standard CUBIC congestion control, with ECN and initial-window configuration.
  • Separate congestion-marking budgets accommodate local datacenter traffic and longer-RTT traffic sharing the same infrastructure.

Results

  • Short-message tail latency comparable to Homa. Across the principal workload and load comparisons, PFQD-DC’s short-message 99th-percentile latency ranged from 0.47 to 1.19 times Homa’s, using the median of five simulation seeds.
  • Up to 53% lower short-message tail latency. In a typical workload, where large messages carry nearly all the bytes, short-message tail latency was 40–53% lower than Homa’s.
  • Faster mid-size messages in the tested high-load case. Messages of 51–255 KB achieved lower median and 99th-percentile completion slowdowns than Homa.
  • Lossless synchronized fan-in with CUBIC and ECN. The tested configuration handled simultaneous responses from up to 15 senders without packet loss.
0 129 258 387 517 Effective Mbps CHAT (100KB) WEB (1MB) STREAM (10MB) DOWNLOAD (100MB) NETFLIX (500MB) 1.3× 1.3× 2.1× 2.8× 1.7× CUBIC BBRv3 WarpRATE

Experience WarpRATE™

Real results. Real networks. Real impact.

Try AccurateSpeedTest.com