Hacker News Reader

This sort of scaling seems to be logarithmic. 16MB is 5 orders of magnitude greater than 512KB, however the logarithmic difference is 3.46. The same increase in 5 orders of magnitude beyond 16MB is 10.39 (at 524GB) which is considerably more impressive.

That's some fancy math and it probably is completely off, but another explanation is NIC performance doesn't follow exponential scaling for unexplained reasons.

View on HN

The article suggests interrupt load is the major issue, although doesn't really say enough to inform us. A 12.5ms buffer means potentially just 80 wakeups/sec. They don't mention whether they use multiple hardware queues or if all interrupts hit a single core.

I'd also be interested to know if polling mode was tested with any of the cards, and why it didn't work out

We use multiple hardware queues. Some details in this old blog post:

http://blog.cloudflare.com/how-the-cloudflare-team-got-into-...

We're continuing to experiment with polling-based network queues.

Awesome, thanks!

Reply on news.ycombinator.com