4 ms·
How large does a TCP buffer have to be before the cost of running beefier hardware compares with deploying to tons of edge locations instead of a couple of loca
by johncolanduoni 5y ago
How large does a TCP buffer have to be before the cost of running beefier hardware compares with deploying to tons of edge locations instead of a couple of locations on each continent? At 300ms of latency you only need 37 megabytes per 1Gbps of link capacity. I can’t imagine putting a 100Gbps link on a machine that couldn’t spare 3.7 GB of RAM without breaking a sweat. TCP window scaling (and SACK) exists for a reason.
- svara 5y agoThat's a good point about window scaling solving the problem in theory. Still, empirically, latency has a huge impact on throughput. Part of the reason might be that many transfers are actually too small to take full advantage of window scaling? Maybe there are other factors at play as well that I'm not aware of. Check out for example this speed measurement tool from AWS [0], which shows a really drastic influence of physical distance on transfer speeds (particularly for the small files where I get >10x differences, but even at 5 MB I see >3x differences between a close-by zone and those roughly on the other side of the planet). That happens even with their "global accelerator" service which gives them a lot of control over the transfer parameters. [0] https://speedtest.globalaccelerator.aws/#/ https://speedtest.globalaccelerator.aws/#/