CsvWriter performance optimization

[snaumenko-st]

I noticed that CsvWriter allocated huge amount of memory during serialization, much more than the size of the data, up to 3 orders of magnitude. It doesn't have memory leaks in my scenario, however such memory traffic creates GC pressure and has noticeable performance impact on my application.

There are some breaking changes to the public types though. But they can be easily fixed if needed. I'd also consider to make them internal in further versions to lower the probability of someone uses them directly.

So that's what is implemented here:

• caching of writer action for collections. We can optimize for the case when most of objects in the collection have the same type.
• removed duplicate type resolving. GetType() is called too often.
• optimized Type HashCodes. No need to get AssemblyQualifiedName to get it's HashCode, we can get it directly from the Type. However both approaches don't guarantee no collisions.
• reduced memory allocation during writes.

Before

Method	Typed	Mean	Error	StdDev	Gen0	Allocated
WriteObject	False	561.6 ms	10.92 ms	14.58 ms	88000.0000	1059.73 MB
WriteStruct	False	545.8 ms	10.69 ms	15.67 ms	91000.0000	1090.25 MB
WritePrimitive	False	441.7 ms	8.80 ms	10.47 ms	77000.0000	922.42 MB
WriteExpando	False	992.4 ms	19.65 ms	27.54 ms	100000.0000	1204.7 MB
WriteDynamic	False	853.7 ms	16.69 ms	19.87 ms	100000.0000	1204.7 MB
WriteObject	True	360.8 ms	7.13 ms	7.33 ms	31000.0000	380.7 MB
WriteStruct	True	354.3 ms	7.08 ms	10.15 ms	31000.0000	380.7 MB
WritePrimitive	True	249.8 ms	4.75 ms	4.67 ms	23333.3333	281.52 MB
WriteExpando	True	984.7 ms	18.81 ms	15.71 ms	100000.0000	1204.7 MB
WriteDynamic	True	858.9 ms	16.42 ms	16.86 ms	100000.0000	1204.7 MB

After

Method	Typed	Mean	Error	StdDev	Gen0	Allocated
WriteObject	False	126.46 ms	0.599 ms	0.500 ms	10000.0000	121.3 MB
WriteStruct	False	135.19 ms	1.728 ms	1.532 ms	12500.0000	151.82 MB
WritePrimitive	False	64.79 ms	1.078 ms	0.955 ms	5000.0000	60.26 MB
WriteExpando	False	531.63 ms	4.189 ms	3.918 ms	23000.0000	281.52 MB
WriteDynamic	False	399.09 ms	6.453 ms	5.720 ms	22000.0000	266.26 MB
WriteObject	True	129.50 ms	1.278 ms	1.067 ms	10000.0000	121.3 MB
WriteStruct	True	122.63 ms	1.732 ms	1.620 ms	10000.0000	121.3 MB
WritePrimitive	True	62.77 ms	0.978 ms	0.867 ms	5000.0000	60.26 MB
WriteExpando	True	589.30 ms	6.714 ms	6.280 ms	23000.0000	281.52 MB
WriteDynamic	True	464.30 ms	3.689 ms	2.880 ms	22000.0000	266.26 MB

[snaumenko-st]

Eliminating all remaining string allocations seems to me much more difficult. As almost half of the code needs to be rewritten from scratch. And of course we don't have convenient support for that in earlier versions of .net. So this PR is mostly a compromise.

[JoshClose]

I'm working on a new parser and serializer that uses spans and has 0 allocations, outside the initial buffer. I'm not sure how I'm going to fit it into the existing codebase yet.

[botinko]

@JoshClose Are there any chances that this PR will be concidered / reviewd / merged and then published in near future? We would really like to fix a performance issue in our project that is fixed in this PR. If not, then we would consider creating a fork.

[JoshClose]

I'll try and get to it soon.

[JoshClose]

Does this same issue happen when reading?

[snaumenko-st]

Does this same issue happen when reading?

No, when reading it's much better, and It allocates only a few bytes per operation.

[JoshClose]

This has been merged. I made other changes before you put in your second commit, that aren't in there, but I had done the same thing myself. Please look over everything and try it out. I will create a release once you do.

[snaumenko-st]

@JoshClose I've tried the same benchmarks, and the results are almost the same, except with ExpandoObject. However it makes no sense to us. Thanks a lot!

P.S. My CsvReader benchmarks before and after your latest fix show almost the same results. So, I'm not sure whether it improved anything.

[JoshClose]

ExpandoObject uses IDynamicMetaObjectProvider so I just removed the ExpandoObject specific code. Apparently it was faster. I need to do some work on the IDynamicMetaObjectProvider code to make it faster.

My CsvReader benchmarks before and after your latest fix show almost the same results. So, I'm not sure whether it improved anything.

Maybe not, but there is definitely less computations. Plus, it's consistent with the writer now.

I'll push a new build soon. Just need to update the change log.

[JoshClose]

I changed expando to use any IDictionary<string, object> now, so that should cover most dynamics. It'll fall back to IDynamicMetaObjectProvider, which is slower.

I added a fast dynamic object that is created instead of ExpandoObject when reading.

[JoshClose]

This will be in the next release 32.0.0 on NuGet.