StorageValue

[Scooletz]

This PR introduces a notion of StorageValue that represents a non-trimmed 32-byte long value persisted contracts' storage. The main goal is to remove the clutter of new byte[] that was used to keep the values and track the changes for the storage, and replace it with a simple struct. To make it work a storage is required to ammortize all data structures that previously hold a ref to byte[] and now need to keep the StorageValue. The storage for it is called StorageValueMap and allows mapping an arbitrary value of StorageValue to a wrapped Ptr struct that is a size of a ref/ptr.

The pointer is used in the storage cache, so that ConcurrentDictionary nodes should be smaller now (value is 8 bytes instead of 32) and should use less memory. They also directly return a platform size value so less copying will occur.

The actual size of the amendment in VirtualMachine is small, as only SLOAD/SSTORE TLOAD/TSTORE are affected. Additionally the tracing seam has been strongly amended.

One more area that is touched is a different implementation of .WithoutLeadingZeros() that is ~ 1.5 times faster that the
Bytes.WithoutLeadingZeros(). It uses the fact that we know the length and for 32 bytes it's easy to vectorize it.

Additional cost of this approach is 36MB of memory that is never released, cleared or GCed.

Benchmarks

A set of benchmarks, using storage related methods: reset, reading with block cache warmed up and a cold cache. Additional set is added showing that the mapping penalty needs to be paid somewhere.

The benchmarks use .IsZero for making some operation on the retrieved value. The new IsZero is quite fast as all it does is

; Nethermind.Core.StorageValue.get_IsZero()
       vmovups   ymm0,[rcx]
       vptest    ymm0,ymm0
       sete      al
       movzx     eax,al
       vzeroupper
       ret
; Total bytes of code 19

Micro-benchmarks

Before

Method	Mean	Error	StdDev	Gen0	Code Size	Allocated
Just_reset	11.38 ns	0.160 ns	0.149 ns	-	1,410 B	-
PreCached_small_indexes	119.46 ns	1.379 ns	1.152 ns	0.0095	6,310 B	122 B
NotCached_small_indexes	822.13 ns	10.680 ns	9.990 ns	0.1078	4,642 B	1354 B
Set	70.04 ns	0.400 ns	0.354 ns	0.0057	1,840 B	72 B

After

Method	Mean	Error	StdDev	Gen0	Code Size	Allocated
Just_reset	11.16 ns	0.079 ns	0.070 ns	-	1,434 B	-
PreCached_small_indexes	81.34 ns	0.811 ns	0.758 ns	0.0082	6,083 B	104 B
NotCached_small_indexes	641.38 ns	3.634 ns	3.034 ns	0.0830	4,468 B	1050 B
Set	75.41 ns	0.722 ns	0.676 ns	0.0057	2,342 B	72 B

CoarseGrained

Before

Method	Bytecode	Mean	Error	StdDev	Gen0	Gen1	Allocated
Execute	Byte[14]	4.257 us	0.0385 us	0.0360 us	0.2823	-	3.46 KB
Execute	Byte[27]	7.250 us	0.0678 us	0.0634 us	0.5569	0.0076	6.87 KB
Execute	Byte[36]	2.883 us	0.0120 us	0.0112 us	0.1488	-	1.84 KB

After

Method	Bytecode	Mean	Error	StdDev	Gen0	Gen1	Allocated
Execute	Byte[14]	3.960 us	0.0248 us	0.0220 us	0.2823	-	3.46 KB
Execute	Byte[27]	6.938 us	0.0735 us	0.0652 us	0.5493	0.0076	6.74 KB
Execute	Byte[36]	2.671 us	0.0237 us	0.0221 us	0.1373	-	1.71 KB

Running nodes

Comparison using GitHub action for RPC (comparable hardware)

Changes

• removal of Span<byte> and byte[] based code for storage
• introduction of a readonly struct StorageValue and its pointy counterpart Ptr
• introduction of ref returning and ref accepting API for storage related operations
• amendments in VirtualMachine in storage management opcodes, including gas calculations

Types of changes

What types of changes does your code introduce?

• Bugfix (a non-breaking change that fixes an issue)
• New feature (a non-breaking change that adds functionality)
• Breaking change (a change that causes existing functionality not to work as expected)
• Optimization
• Refactoring
• Documentation update
• Build-related changes
• Other: Description

Testing

Requires testing

• Yes
• No

If yes, did you write tests?

• Yes
• No

Notes on testing

Compare on various sizes of VMs

Documentation

Requires documentation update

• Yes
• No

Requires explanation in Release Notes

• Yes
• No

[Scooletz]

A bit related to #8442 although working on a different level.

[Scooletz]

Let's pause reviews for a while. Using more comparable machines showed different results from the ones presented before, showing that this branch is slower.