[94] in Kakapo Windows Team
status update on slow startup times and afs connections
daemon@ATHENA.MIT.EDU (Thomas L. Thornton)
Thu Sep 18 12:05:28 2003
Date: Thu, 18 Sep 2003 12:05:26 -0400 (EDT)
Message-Id: <200309181605.h8IG5QOE019032@the-rim.mit.edu>
From: "Thomas L. Thornton" <tomt@MIT.EDU>
To: contact-container-admins@MIT.EDU
CC: kakapo@MIT.EDU, ferrara@MIT.EDU
This follows up on a thread "AFS &/or XP issues" we started almost a
month ago. Two issues of interest to the win.mit.edu container
administrator continue to exist and deserve this report:
1. Some XP machines spend ten or more minutes during reboot in the
"Applying computer settings" state;
2. Sometimes AFS takes several tries before successfully restarting
the service.
There is much mail about these misbehaviors, and Casetracker numbers
probably associated with the above, respectively, are:
1. 404542, 408880
2. 410179, 411090, 417061, 418345, 418352, 426296, 428304
In August we noticed an erroneous service-restart setting missing from
installer, fixed it in our code base and we expect this problem to
disappear upon our next AFS release. Over the past three weeks we
have found a couple more possible reasons for service failure. We
have delayed packaging these fixes expecting to address the long
reboot time problem.
The long reboot problem has been particularly burdensome. Paul asked
MS PSS about this weeks ago. They got back to us last week, saying
that the slowdown can be addressed by adding netman to AFS service
dependencies in the registry. Alas, testing this indicates no
improvement.
Meanwhile we brainstormed on possible workarounds like reordering the
listing of network providers, disabling the "Network Connections"
service, looking for interactions with WebDAV or the WebClient
service, the ShellSvcGroup and the GP that sets "Always wait for the
network at computer startup and logon." No luck with any of these.
In June by Microsoft advice, we rewrote the code for determining which
LAN adapter to use. The old code uses an undocumented API. According
to Microsoft, this API should not be used, and will go away in the
future. However, their suggested alternative is unimplemented on W2K,
so for the currently-deployed AFS we made two code paths and expected
XP and later OSs to work via the documented and supported API.
As a last resort Paul and Joe built a service that uses only the old
API. Testing the old API service on XP machines eliminated the boot
delay. The service seems to function normally (e.g.users' tokens are
acquired, the home directory is available, etc.) and with machine
startup times comparable to W2K.
Therefore, now we *may* have a workaround, with two big
qualifications. First caveat, this must be temporary - the service
uses the undocumented API that may break with any future SP or OS.
Second, it has been tested on only IS machines and a couple in FSS, so
needs some more field testing before we package and deploy it.
Chad Dupuis, Dan Stratila, Phil Thompson, could you volunteer a test
on each of your slowest-booting machines? Please let me know.
-Tom
