Investigate high exclusive time in MSD.GetNode

[rynowak]

This is sampling data from the LargeJsonApi app using the POST test scenario:

The total time spent processing the request was 28s, and 3780ms of exclusive time (own time) was spent in ModelStateDictionary.GetNode(...). This is 13.5% of the app's processing time. https://github.com/aspnet/Mvc/blob/dev/src/Microsoft.AspNetCore.Mvc.Abstractions/ModelBinding/ModelStateDictionary.cs#L509

You can see from the trace that new StringSegment and ModelStateNode.GetNode were not inlined, so I immediately suspect string.IndexOfAny which is implemented natively by the runtime. Our search space is generally constrained to dot-separate property names, which means the search text is small, and the number of delimiters we're looking for is 2. This seems like a good candidate for hand-optimizing.

[dougbu]

Our search space is generally constrained to dot-separate property names

Need to handle look-ups for random variables, collection elements, and random mixes of supported expression types e.g. @Html.TextBoxFor(m => variable1[variable2].Property.SubProperty).

[benaadams]

Inlines into 060000F9 ModelStateDictionary:GetNode(ref,bool):ref:this
  [0 IL=0056 TR=000058 06000018] [FAILED: unprofitable inline] StringSegment:.ctor(ref,int,int):this
  [0 IL=0065 TR=000066 060002EF] [FAILED: unprofitable inline] ModelStateNode:GetNode(struct,bool):ref:this
  [1 IL=0087 TR=000021 06001382] [profitable inline] String:IndexOfAny(ref,int):int:this
  [0 IL=0144 TR=000167 06000018] [FAILED: unprofitable inline] StringSegment:.ctor(ref,int,int):this
  [0 IL=0153 TR=000175 060002EF] [FAILED: unprofitable inline] ModelStateNode:GetNode(struct,bool):ref:this
  [2 IL=0163 TR=000129 060002E8] [below ALWAYS_INLINE size] ModelStateNode:get_Key():ref:this
  [3 IL=0172 TR=000139 060002E9] [below ALWAYS_INLINE size] ModelStateNode:set_Key(ref):this

[benaadams]

Made the new StringSegment inline for ya dotnet/extensions#159

[benaadams]

Also string.IndexOfAny initialises a probabilistic map before it begins searching

https://github.com/dotnet/coreclr/blob/master/src/classlibnative/bcltype/stringnative.cpp#L485-L537

If your Delimiters are fairly constant you could extract that out and use it multiple times; or even Vectorize the test...

[rynowak]

WOW thanks for digging into this. We're looking into relatively small strings (~6-100 characters) for 2 possible delimiters. I'm wondering how a naive implementation would compare. This method is called often, but the inputs are not complex.

[benaadams]

Have something for this... got very distracted on it O_o