IIS 7 C++ Module API Sample: Controlling http.sys caching

Introduction

In my previous post we saw that http.sys caching can be very useful for high traffic (gigabits/sec) sites. In this post we create a sample module that allows configuring which URLs can enter the cache.

Code

REQUEST_NOTIFICATION_STATUS
MODULE::OnSendResponse(
    IHttpContext*           pContext,
    ISendResponseProvider*  pProvider
)
{
    if( !pProvider->GetHeadersBeingSent( ) )
    {
        //
        // headers not being sent, not first send response notification, nothing to do
        //
        goto Finished;
    }

    if( !g_pUrlHash->Exists( pContext->GetScriptName( ) ) )
    {
        //
        // not a pre-determined "hot" url, no caching..
        //
        IHttpResponse* pResponse = pContext->GetResponse( );
        pResponse->DisableKernelCache( 123 );
    }

Finished:

    return RQ_NOTIFICATION_CONTINUE;
}

• RQ_SEND_RESPONSE is a "non-deterministic" notification that can fire multiple times (e.g. whenever someone calls IHttpResponse::Flush). To determine the first time we're being notified, the module checks ISendResponseProvider::GetHeadersBeingSent. This is only true on the first notification for the request.
• IHttpContext::GetScriptName is often the URL you want to compare to. It has been "sanitized" and had PATH_INFO removed.
• IHttpResponse::DisableKernelCache trumps any cache policy configured with IHttpResponse::GetRawHttpResponse()->...

Setup

• disabled my file cache module (simply removed it from the pipline - "appcmd uninstall module FileCacheModule").
• disabled the "hotness" detection algorithm in iiscore ("appcmd set config -section:serverRuntime -frequentHitThreshold:1")

Test

This time I'm using a newer dual-core machine with 1GB ram and running the perf client locally. There are 4000x184kb "hot" files, and 4000x184kb "cold" files. Total = 8000x184kb = 1.4GB. Hot files are hit 32x more frequently than cold files. For the Cacheit test, I installed the module and only gave it the full list of 4000 hot URLs.

Request/sec came from wcat's log. Cache contents was determined by running "netsh http show cachestate > out.txt" and then "findstr /i hot out.txt" and "findstr /i cold out.txt".

Conclusion