Improve Pickling performance

[adamsitnik]

This PR is similar to #64 so again we have two structs that implement writing to Array and Stream. By using them as generics all virtual method calls are eliminated which allows for inlining and other optimizations.

Difference:

• stream.WriteByte(v) is just array[position++] = v.
• Direct write to the underlying memory buffer with no extra copy step:

WAS:
BinaryPrimitives.Write(tempBuffer, value);
stream.Write(tempBuffer, 0, tempBuffer.Length);
IS:
BinaryPrimitives.Write(outputBuffer, value);

• Serialize strings as Utf8 with no memory allocation (this is going to help a LOT in the Q13 TPC-H benchmark):

WAS:
byte[] encoded = Encoding.UTF8.GetBytes(str); // array allocation
stream.Write(encoded, 0, encoded.Length);
IS:
Encoding.UTF8.GetBytes(source, str.Length, destination + position, byteCount);

• more inlining
• faster type checks
• hot path for strings (1 extra method)
• possibility to serialize to provided byte buffer with no managed allocations

Benchmark Results:

Following benchmark results are for pickler with useMemo=false (dotnet/spark use case) and 1000 elements in each array.

Serializing an array of given type to array of bytes:

Type	Mean WAS	MEAN IS	WAS\IS
bool[]	51.83 us	45.491 us	1.14
double[]	27.41 us	9.192 us	2.98
int[]	14.44 us	6.362 us	2.27
string[]	77.55 us	27.774 us	2.79

Serializing an array of objects with boxed values inside (dotnet/spark scenario, every array can contain values of different types)

Type	Mean WAS	MEAN IS	WAS\IS
bool[]	22.96 us	16.951 us	1.35
double[]	43.37 us	22.033 us	1.97
int[]	31.94 us	20.530 us	1.56
string[]	81.45 us	43.355 us	1.88

Serializing an array of given type to provided array of bytes:

I have added a new method that accepts a byte buffer by reference and can resize it if needed. This will be helpfull in dotnet/spark scenario where input comes in very similar batches and output is almost always of the same size (for example: an array of 100 doubles, an array of 100 bools) .
The goal is to allocate an array of bytes once and then write to it as many times as we want with almost no managed memory allocations (72 bytes are going to be allocated, independent from input size).

Type	Mean WAS	MEAN IS	WAS\IS	Alloc WAS	Alloc IS
bool[]	51.83 us	44.662 us	1.16	34.17 KB	32097 B
double[]	27.41 us	5.862 us	4.68	40.82 KB	72 B
int[]	14.44 us	5.290 us	2.73	10.66 KB	72 B
string[]	77.55 us	25.636 us	3.03	54.95 KB	72 B

Serializing an object[] to provided array of bytes

Type	Mean WAS	MEAN IS	WAS\IS	Alloc WAS	Alloc IS
bool[]	22.96 us	17.184 us	1.34	2.9 KB	72 B
double[]	43.37 us	18.474 us	2.35	40.81 KB	72 B
int[]	31.94 us	18.824 us	1.70	10.65 KB	72 B
string[]	81.45 us	37.711 us	2.16	54.95 KB	72 B

Serializing to Stream

No regressions, minor improvements:

Type	Mean WAS	Mean IS
bool[]	51.45 us	49.901 us
double[]	27.74 us	27.089 us
int[]	15.36 us	15.150 us
string[]	77.18 us	60.875 us

Boxed input (object[])

Type	Mean WAS	Mean IS
bool[]	22.14 us	19.775 us
double[]	41.35 us	39.768 us
int[]	31.21 us	28.752 us
string[]	79.40 us	72.461 us

/cc @danmosemsft @stephentoub, @imback82, @eerhardt, @rapoth

[irmen]

@adamsitnik random question (excuse my lack of understanding of how Spark and .NET Spark work): are you only using Pyrolite's pickle protocol implementation to talk to Spark? I mean; Pickle is a Python serialization protocol and Spark is primarily a Java system.

[imback82]

@adamsitnik random question (excuse my lack of understanding of how Spark and .NET Spark work): are you only using Pyrolite's pickle protocol implementation to talk to Spark?

Yes. dotnet/spark is piggybacking on how PySpark works. So, it needs to follow the same protocol as the Python worker, which uses pickling for Ser/De to transfer data between the Python worker process and Spark JVM executor process.

[eerhardt]

Looks good. Awesome work @adamsitnik!

[irmen]

Amazed by your improvements.

I wonder if similar improvements are possible for the Java implementation.

[irmen]

@adamsitnik @stephentoub I have just pushed a new release (4.26) to nuget.org

[adamsitnik]

I have just pushed a new release (4.26) to nuget.org

@irmen awesome, thank you!!