Stsadm / new SPSite is slow

3 minute read

Update: Get the script here

The Story

A couple of days ago, developer extraordinaire Itay Shakury was doing performance tuning on one of our SharePoint applications, and came to me with a problem - creating a new SPSite object took about 30 secs. The stranger part was that only the first creation of SPSite in every appDomain is slow.
We’ve tried the usual things:

  • Using SQL Profiler to check for DB-delays. Nope, all queries were as quick as a fox
  • Perfmoning the servers to death, looking for CPU/memory/disk spikes. No luck again

By the way, we used this bit of code to check exactly how slow the SPSite creation is:

function Log-Message($Message) {[DateTime]::Now.ToString('[hh:mm:ss.fff]')+" $Message"}
$Url = 'http://SERVER/Url';
Log-Message 'Loading Assembly'
[Reflection.Assembly]::LoadWithPartialName('Microsoft.Sharepoint') | Out-Null;
Log-Message 'Creating SPSite'
$Site = new-object Microsoft.Sharepoint.SPSite $Url
Log-Message 'Finished'

Now I did some thinkning. What happens only one time in an AppDomain? I thought of 3 things:

  • Poorly written singletons. If that were the case, people all around the world would be having problems, and not just us.
  • Connections to DBs. Using SQL Profiler we found out that the connection to the ConfigDB only happens in the last of the 30 seconds, meaning the DB is not to blame…
  • Lazy loading. (Spoiler: This is it) For those of us who forgot what it is, lazy loading is “don’t look for / load dlls until the first time they’re actually needed”. How can we check that

I suddenly remembered some problem I hadn’t run into myself, but is famous in my team:
SQL Server Management Studio was slow to load (KB918644). It happened because the studio’s assemblies (dlls) were digitally signed by MS using Authenticode, that uses some sort of SSL-like certificates, and the studio was trying to check MS’s CRL (Certificate Revocation List) to make sure the digisignature isn’t revoked. Because we work in a disconnected environment, it couldn’t succeed.
I’ve decided to NetMon the sever, and sure enough, my script was DNS-querying for crl.microsoft.com, and obviously failing.
After solving this issue, as a bonus, my Stsadm.exe commands were WAY faster. Until now, it would take the process about 1 min to tell me I have the wrong input, and I would die a little bit inside.

Possible Solutions

1. Allow access to crl.microsoft.com

If you’re lucky(?) enough to be working in a blocked, not disconnected environment, consider simply allowing traffic to the crl site through your proxy/firewall/whatever.

2. Stop checking for CRLs

If you ouldn’t care less if someone stole MS’s certificate and created an evil harepoint.dll just kidding. We all fear that), you can tell Windows to stop checking for CRLs altogether.
The user-specific way:
Note: This setting affects all software run by that user. You shouldn’t use turn it off if this user on this server is used to process smartcard logons, for example.
Either turn off Inetcpl.cpl > Advanced > Security > Check for publisher's certificate revocation Or set this regkey:

HKCU\Software\Microsoft\Windows\CurrentVersion\WinTrust\Trust Providers\Software Publishing:State=00023e00

Another note: Iv’e seen people scripting something to change the settings for all keys under HKEY_USERS. While this seems ok, it will actually change settings for all new users (through .DEFAULT) and for all existing users with their registry hive currently loaded. Not all users have their hive loaded at every moment! If you’re into automation, you’re better off with using Group Policy Preferences (use user preferences, and server-name-based targeting). The machine / app specific way:
Use the machine/app/web.config to disable generation of publisher evidence, the CRL-related process, thusly:

<?xml version="1.0" encoding="utf-8"?>
<configuration>
      <runtime>
              <generatePublisherEvidence enabled="false"/>
      </runtime>
</configuration>

(Of course you don’t blank your .config files, only add this setting) Update (27/12): Apparently, on x64 machines, you should change boththe x86 and the x64 .config files.
Update (28/1/14): I uploaded the script we used

3. Fetch the CRLs yourself

If the CRLs stored in the server’s certificate store are fresh enough, it shouldn’t dial home for new ones.
Download these:

Add to certificate store:

certutil -addstore CA CodeSignPCA.crl   
certutil -addstore CA CodeSignPCA2.crl

Could be distributed through Group Policy as well, but I didn’t try it.

4. Mess with the process

I don’t like this option, but If you’re feeling malicious, you can just use your hosts file to point crl.microsoft.com into one of your servers (e.g. 127.0.0.1), causing windows to skip the 15-sec timeout issue and fail immediatly, when the server will refuse to hand him the CRLs.

Credit

Kudos for Dirk Van den Berghe for handing out the methods for solving the issue.