.NET Framework 4 is our first release since we shipped FX 3.5 SP1 (FX 4 beta 1 is now available here: http://msdn.microsoft.com/en-us/vstudio/dd582936.aspx). FX 3.5 SP1 contained major changes to NGen – features that improved startup performance, security, NGen time and compilation working set – described at length in this MSDN article: http://msdn.microsoft.com/en-us/magazine/dd569747.aspx.

 

In FX 4 we shifted our primary focus from startup performance to framework deployment. As many of you are aware, the performance win from pre-compiling your application using NGen comes at a cost – the time taken to generate the NGen images on the end-user machine.  In .NET 4 we’ve lowered that cost substantially in two ways.  First, we’ve made NGEN multiproc-aware.  In many cases, your assemblies (and ours) will now NGEN about twice as fast!  In addition, we’ve substantially reduced the number of situations in which we need to regenerate NGen images.  Our new Targeted Patching feature means that for many .NET Framework patches, we can now modify just the affected assemblies and not have to re-NGEN any other dependent assemblies.  Combined, these two features mean you and your users will spend a lot less time running NGEN.  Best of all, you don’t have to do anything to get these benefits – they’ll happen automatically when you use .NET 4. 

We thought you may want to learn a little more about how these features work and the situations in which you’ll get these benefits.  In addition, since .NET 4 is a SxS release, there is a bit of complexity under the covers to make sure assemblies get NGen-ed against the matching runtime.  In general things will just work the way you would guess or expect them to, but below, I’ll go into some more detail about exactly what happens in various interesting cases.   

If you have any comments or questions about any of these, we’d love to hear from you. Most of these features are available in the FX 4 beta 1 build.

·        NGen SxS: Making NGen work correctly when 2 major versions of the CLR are installed side-by-side

·        Multi-proc NGen: Enabling use of multiple cores/processors during NGen to make it faster

·        Targeted Patching: First step towards making NGen images less “fragile” – avoids having to recompile all managed assemblies when a FX dependency is patched

·        No NGen in partial trust: Deprecated support for generating and loading NGen images in partial trust applications

NGen SxS

To a large extent FX 4 is the first SxS release for NGen (the NGen infrastructure in FX 1.0 and 1.1 was rudimentary). We’ve now architected NGen such that the 4.0 ngen.exe tool can be used to generate NGen images for both 2.0 and 4.0 MSIL assemblies without having to pass any special arguments. NGen uses the same logic that is used at application runtime (this logic resides in the shim) to determine the CLR version to load and run an application against. Thus,

%WINDIR%\Microsoft.NET\Framework\v4.0.xxxxx\ngen.exe install <4.0 assembly> will generate an NGen image compiled against the 4.0 runtime.

%WINDIR%\Microsoft.NET\Framework\v4.0.xxxxx\ngen.exe install <2.0 assembly> will generate an NGen image compiled against the 2.0 runtime (if .NET Framework 2.0/3.0/3.5 is installed on the machine. Note that applications built against .NET Framework 2.0 won’t run against .NET Framework 4.0 by default, and thus by default we don’t NGen 2.0 assemblies unless CLR 2 is installed).

%WINDIR%\Microsoft.NET\Framework\v4.0.xxxxx\ngen.exe install <2.0 assembly> /ExeConfig:<Path to a 4.0 EXE> will generate an NGen image compiled against the 4.0 runtime.

%WINDIR%\Microsoft.NET\Framework\v4.0.xxxxx\ngen.exe install <2.0 EXE with a config file that indicates 4.0 as the preferred runtime> will generate an NGen image compiled against the 4.0 runtime.

%WINDIR%\Microsoft.NET\Framework\v4.0.xxxxx\ngen.exe install <2.0 assembly> /ExeConfig:<Path to a 2.0 EXE with a config file that indicates 4.0 as the preferred runtime> will generate an NGen image compiled against the 4.0 runtime.

There was also work required to make the 2.0 NGen Service (clr_optimization_v2.0.50727_32|64) work SxS with the 4.0 NGen Service (clr_optimization_v4.0.xxxxx_32|64). Only one service (the latest) is active at any time – installing FX 4 disables the 2.0 service and it is re-enabled if/when FX 4 is uninstalled.

Multiproc NGen

NGen is now aware of multiple processors/cores and can compile up to 6 assemblies in parallel (the parallelism is at the level of assemblies). To avoid impacting foreground activities, NGen only runs in this aggressive mode when assemblies are being compiled synchronously i.e. the NGen Service still runs on one processor. Synchronous NGen commands (such as ngen.exe install <assembly>, ngen.exe update, ngen.exe ExecuteQueuedItems) will now use multiple processors/cores whenever possible (we also factor in amount of RAM when determining how many cores/processors to use). Since the parallelism is at the level of assemblies, the most effective way to enable use of multiple processors for NGen is to do the following:

ngen.exe install /queue:1 <MyImportantAssembly#1>

ngen.exe install /queue:1 <MyImportantAssembly#2>

ngen.exe install /queue:1 <MyImportantAssembly#N>

ngen.exe install /queue:3 <MyAssembly#N+1>

….

ngen.exe install /queue:3 <MyAssembly#M>

ngen.exe ExecuteQueuedItems 1 //Synchronously compiles all important (priority 1) assemblies during set up using multiple cores whenever possible; other (priority 3) assemblies will be compiled in the background by the NGen Service at machine idle-time.

As part of this work we reworked FX 4 set up to use the NGen pattern above.

Targeted Patching

NGen images thus far have been nothing more than “cached JIT-compiled code and CLR data structures” – as a consequence they’re completely fragile; any change to the underlying CLR or to any managed dependency invalidates them and requires them to be regenerated. For example, any change to the CLR or to basic assemblies like mscorlib and System that all/most assemblies depend upon invalidates all NGen images installed on the machine. Since a machine could have several hundreds of NGen images, the cost of regenerating them after FX servicing events is high. In CLR 4 we took a first step towards making NGen images less fragile to avoid the cascaded cost associated with .NET Framework updates. In particular, FX updates that only involve fixes to bodies of existing methods (that aren’t generic, and aren’t inlined across NGen image boundaries) will no longer require recompiling dependent NGen images (the old NGen images can be used with the new serviced dependency). Although this may sound trivial (and beg the question why the system didn’t have this attribute to begin with J), since NGen images had never been architected to be anything but serialized JIT-ed code and CLR data, accomplishing this turned out to be major feat. From reworking hardbinding, doing major performance work to recover the perf impact, to writing an IL post-processing tool that normalizes metadata tokens, detects Targeted Patching-compatible vs. incompatible changes, and flags compatible changes such that existing NGen images can rewired to the updated dependency, plugging that tool into our build system, and revising NGen cross-module inlining rules, this is a major effort that several of us have been working on for several months.

Some of the work for this (such as the changes to hardbinding and corresponding the performance work) are included in the beta 1 build, but this feature will really come online once we start servicing FX 4. You can find out more in the Channel 9 video on Targeted Patching.

NGen in partial trust

In FX 2 NGen images can be generated by running commands such as ngen.exe install <Path to assembly on intranet share> and the generated image loaded in applications running in partial trust. We believe NGen images aren’t used in partial trust applications very much and our current model was broken, so we’ve disabled loading of NGen images in partial trust in FX 4. In the future (post CLR 4) we intend to rework this as part of simplifying the overall NGen story (i.e. change the model where the only way to generate NGen images is to issue commands from an elevated command prompt). If you’re using NGen in a partial trust application today, we’d like to hear from you!

 

Surupa Biswas

CLR Codegen Team