3 ms·
Nice work. I guess it'd just as quick if it sent up to ~4kB - that's the normal TCP initial window? https://tools.ietf.org/html/rfc3390 https://tools.ietf.org/h
by mkj 9y ago
Nice work. I guess it'd just as quick if it sent up to ~4kB - that's the normal TCP initial window? https://tools.ietf.org/html/rfc3390 https://tools.ietf.org/html/rfc3390
- unmole 9y agoAt least in Linux, ICWND is set to 10 now: https://tools.ietf.org/html/rfc6928 https://tools.ietf.org/html/rfc6928
- mkj 9y agoHm. I what the distribution of latency vs bandwidth is like, whether 14kB or 2 round trips takes longer for most people?
- Dylan16807 9y agoProbably the two round trips. A semi-nearby server might be 50ms per round trip, and in that case any connection beating 1mbps can transfer 14KB in less than two round trips. A server could easily be 150-250ms away and make those round trips far worse.
- londons_explore 9y agoIt also depends on the buffer size. Some routers might only buffer 1 millisecond of line-rate data or less. If that's the case, that initial congestion window of 10 could overflow the buffer, causing loss and a much more costly double-round-trip to resend.
- Dylan16807 9y ago1 millisecond of buffering is geat on a high speed core link. But if you apply that to a 5 megabit connection you've basically decided you'll only buffer one packet. A slow-speed bottleneck link needs to have more capacity. The charts in http://www.nctatechnicalpapers.com/Paper/2011/2011-right-sizing-cable-modem-buffers-for-maximum-application-performance/download http://www.nctatechnicalpapers.com/Paper/2011/2011-right-siz... suggest that 64KB is often a good buffer size for end links, and that can absorb 10 packets fine.