This project is dicontinued. Although it was able to be faster than the default shared implementation, it's not faster than Kestrel's pool. More reasoning: https://blog.scooletz.com/2018/08/14/thruster-building-a-not-so-fast-memory-pool/

Thruster is a fast and efficient implementation of a MemoryPool<T>

Thruster is based on (some) understanding on nowadays CPUs and low level APIs provided by .NET. The most important desing principles are:

• padding - expanding an object or an array, to reduce a possibility of using the same CPU cache line by threads being run on different CPUs
• region memory mangagement - leasing chunks of memory from a specific region, which allows use of in-region addressing (offset, region number)
• no locks - for claiming and releasing memory when no resize is needed
• efficient data encoding - bitmasks used for leasing, are designed to use 1 CAS operation for renting (when no collisions) and one, branchless unconditional Interlocked.Add to release

Some single threaded benchmarks run with awesome BenchmarkDotNet. The nesting level represents multiple .Rent operations.

BenchmarkDotNet=v0.11.0, OS=Windows 10.0.16299.547 (1709/FallCreatorsUpdate/Redstone3)
Intel Core i7-6700HQ CPU 2.60GHz (Max: 2.20GHz) (Skylake), 1 CPU, 8 logical and 4 physical cores
Frequency=2531250 Hz, Resolution=395.0617 ns, Timer=TSC
.NET Core SDK=2.1.302
  [Host]     : .NET Core 2.1.2 (CoreCLR 4.6.26628.05, CoreFX 4.6.26629.01), 64bit RyuJIT
  DefaultJob : .NET Core 2.1.2 (CoreCLR 4.6.26628.05, CoreFX 4.6.26629.01), 64bit RyuJIT

Multi threaded app depends on the nesting level even more. If there's no nesting, results are comparable. Any nesting will drastically point towards using Thruster.

Rocket Ship designed by Joy Thomas from The Noun Project