CentOS on Microsoft Azure "cloud"

[woodsmoke]

http://www.wired.com/cloudline/2012/06/centos-azure/

woodsmoke

[GreyGeek]

Did Hell freeze over?

(Actually, Microsoft has been using Linux regularly for years. Their update and dispenser server farm, run by Akima, used hundreds of Linux boxes.)

[bsniadajewski]

Nah, Hell just moved to a more temperate climate.

[SteveRiley]

(Actually, Microsoft has been using Linux regularly for years. Their update and dispenser server farm, run by Akima, used hundreds of Linux boxes.)

Two reasons for this. First, Akamai is everywhere. Second, the vast amount of traffic that Windows Update pushes out would overwhelm traditional IP stacks.

Most stacks implement TCP Reno or one of its close cousins, a congestion avoidance algorithm that relies on detecting packet loss. An alternate means to avoid congestion is to use TCP Vegas, which relies on measuring delay. Reno is a pig; it fills the pipe as much as it can, and then backs off when packets disappear. Vegas is more "fair" in the sense that it prefers to play nicely and try to share the circuit with everyone else. The drawback is exactly what you'd imagine: in a room of 100 people, if 99 are pigs, the one dude who likes to share will get clobbered. So Vegas never really took off. (woodsmoke: a political analogy awaits, I'm sure! )

However, there are ways to improve upon Vegas such that a Reno-obsessed Internet won't obliterate it. Windows doesn't include any form of Vegas, and its stack isn't built to allow for choices of congestion avoidance algorithms. Linux, of course, has no such restriction. A modified Vegas, one that dynamically adapts itself to individual flows, allows for each flow to enjoy pig-like performance while nearly eliminating loss. As you might imagine, the sheer amount of traffic generated by Windows Update would benefit greatly from using an algorithm that can drive packet loss -- and the subsequently required retransmission -- to almost zero.