Simple Patch Lets Amazon's EC2 Linux Network Driver Start ~90x Faster

[Jaxad0127]

That's what puzzles me - with the new polling interval it does poll much more frequently, but it's faster.

If it only polls when it's expecting a response, it should take less time since it will find the response sooner.

[bridgman]

It's probably multiple Send Command, Await Response interactions, not just one. Polling roughly 500 times as often over multiple interactions will add up.

That's what puzzles me - with the new polling interval it does poll much more frequently, but it's faster.

[discordian]

It's probably multiple Send Command, Await Response interactions, not just one. Polling roughly 500 times as often over multiple interactions will add up.

Its rather the question, why a paravirtualized driver needs such a sketchy kind up setup procedure. This should be some shared memory, pretty much ready-to-use from the start, the rest some sort of socket operations.

[Jaxad0127]

Am I the only one wondering how tightening up a 5mS polling loop shaved 170mS off the startup time ?

If the poll itself was taking a long time then it still wouldn't take the new startup time down to 2mS.

It's probably multiple Send Command, Await Response interactions, not just one. Polling roughly 500 times as often over multiple interactions will add up.

[NotMine999]

Time is money.

[theriddick]

While its only 171ms, cumulatively it does add up over time if they continue to resolve issues like this. That's the problem with small hanging fruit problems, often very hard to see the value until you knock a fair few of them off the tree!

[CommunityMember]

While reducing the ethernet driver initialization time is overall goodness, quite honestly most of my EC2 instances take *so* much more time to startup the various applications and services that saving a hundred milliseconds is just noise....

[andyprough]

Biggest waste of 171.1 ms in my life. The trips I could have taken with that extra time, the mountains I could have climbed ...

[bridgman]

Am I the only one wondering how tightening up a 5mS polling loop shaved 170mS off the startup time ?

If the poll itself was taking a long time then it still wouldn't take the new startup time down to 2mS.

[microcode]

Kinda wonder why it takes any time at all to initialize a virtual network driver. Surely the host could just leave some state somewhere to tell it where the ring buffers are, then be done with it. I guess I'll have to look at it to find out what they thought was worth doing that wasn't that.