Summary

  • Renato Ornelas measured OpenX traffic over 34 World Cup match nights. At the 24 June high point, a local cache stopped at 79 Gbps, one PNI reached 2×100G, and 165 Gbps of Google traffic still came into Brazil from abroad.
  • His slides say another PNI site had room. The missing evidence is a peak-window record of which prefixes and sessions could actually use that spare capacity.

One peering site had space. The packets took the long way home anyway. At NANOG 98, Renato Ornelas put 165 Gbps on that gap: Google traffic delivered from outside Brazil during the World Cup peak, even while OpenX had room at its other local PNI site. A network diagram could show both ports as available. The flow record showed which route was actually used.

Ornelas is OpenX's founder and CTO. His original presentation describes five-minute flow samples across 34 match nights from 11 June to 19 July. OpenX is a wholesale edge, with more than 60 downstream ASNs and an estimated one to three million subscribers behind those networks. These are its own measurements, not a map of all Brazilian viewing.

On 24 June, Google traffic at that edge reached 416 Gbps. The embedded cache topped out at 79 Gbps. One Cotia PNI site reached its two 100G links; the other had spare capacity. Yet 165 Gbps arrived from outside the country. The session's origin breakdown calls that about 40% of 409 Gbps in the peak window. The 409 Gbps breakdown and the 416 Gbps observed maximum are different readings, and Ornelas did not flatten them into one magic number.

The event was scheduled. That did not make last month's busy-hour chart a useful capacity plan. Ornelas's deck compares a football-free Wednesday window at 48 Gbps with the match window at 416 Gbps. Across the tournament it records about 12 PB above comparable football-free baselines. On a separate Brazil match, half-time took 128 Gbps off the Google peak in ten minutes. Kickoff, pause and extra time moved the network more abruptly than a monthly average could explain.

The tempting story is that an idle local port should have swallowed the 165 Gbps. The slides do not show the per-site prefix, session and steering state needed to prove that counterfactual. They do show the operational question: who could direct that traffic to the roomier site, under what interconnection terms, and on what timescale? Ornelas supplied the flow trace. The remaining record belongs to the parties that chose the path.

Sources: NANOG 98 session · Ornelas's original slides.