changed load behaviour in internet explorer and chrome since 2015-06-02

Web Performance Discuss Test Results
1

Hi Pat,

we observe a dramatical change in load behaviour of entry pages in internet explorer version 10 and chrome (with limited scope) the last days. We measure with a private wpt instance, multistep fork (see GitHub - iteratec-hh/webpagetest at multistep) on windows 7 (physical machines).

We are automating tests with OpenSpeedMonitor and these are the doc ready times from 2015-06-01 to 2015-06-02:
[attachment=495]

Black graph represents chrome measurements, green and blue ones internet explorer. All three graphs represent location queues with two agents. Our hypothesis is that a windows update or something os related on the machines is responsible for the change of load behaviour. We didn’t change anything in configuration or installation of measuring components in respective time slot. First Machines of black and green graph switched 2015-06-01 12:00, last ones and both of blue graph 2015-06-02.

Waterfall internet explorer before switch:
[attachment=496]
Waterfall internet explorer after switch:
[attachment=497]

Before the switch the network stack waited with initialization of first connection and tls handshake until all the dns preresolves are done. So we have a gap in first (html-) resource of about 700 ms.
After the switch we observe preconnections in addition to the pre resolves and all that happens in parallel to the html request.

Waterfall chrome before switch:
[attachment=498]
Waterfall chrome after switch:
[attachment=499]

In chrome waterfalls we don’t see any pre resolving or pre connections. But we see a gap between html request and all the following requests before the switch. After 2015-06-02 this gap is gone and follow-up requests dns lookups happen in parallel.

Unfortunately we couldn’t find any windows updates or something suspicious in windows event logs in respective time slot :frowning:

Maybe you or somebody here can give an explanation for this?
On webpagetest, Dulles location, IE 10 + 11 provide results without any (visible) pre-resolves or pre-connections:
http://www.webpagetest.org/result/150604_TZ_QF1/1/details/
http://www.webpagetest.org/result/150604_FS_QF6/1/details/

Which windows version is running on these agents? Do they run windows updates automatically?

Best regards, Nils

2

The IE 10/11 pre-resolve issue should be tied to a history cache that I could only clear well on IE 7 (which is what the public instance uses).

As far as the change goes, it looks like it’s all from early work right around the TLS negotiation (though in IE it looks like it might be before that). That was always a gap with IE that was something in the OS so I wouldn’t be surprised if an update fixed it but I have no idea what it was.