Description

I've been evaluating the use of gRPC streaming in aspnet at scale for an IoT scenario. I was a little surprised to find out that a single connection requires ~50KB of memory. Since I am planning to scale to >100,000 connections, I would look to maximize the density of the connections per machine.

Looking deeper at the memory usage, a large chunk (12kb) is taken up by 3 4KB preallocated arrays in the HPackDecoder. I wonder if this size is overly aggressive in the general case w.r.t. memory consumption per connection.

Does anyone know why this size was chosen?

.NET SDK:
 Version:           9.0.200-preview.0.25057.12
 Commit:            cf6faae8d1
 Workload version:  9.0.200-manifests.693890c1
 MSBuild version:   17.13.0-preview-24569-04+8f6b8ad0a

Runtime Environment:
 OS Name:     Windows
 OS Version:  10.0.19045
 OS Platform: Windows
 RID:         win-x64
 Base Path:   C:\Program Files\dotnet\sdk\9.0.200-preview.0.25057.12\

.NET workloads installed:
There are no installed workloads to display.
Configured to use loose manifests when installing new manifests.

Host:
  Version:      9.0.0
  Architecture: x64
  Commit:       9d5a6a9aa4

Regression?

No, this allocation is years old.

Data

Analysis

This is the location of the array allocations. If perhaps

https://github.com/dotnet/runtime/blob/0f0f8ce751134cccf11e47c90f781597039146ef/src/libraries/Common/src/System/Net/Http/aspnetcore/Http2/Hpack/HPackDecoder.cs#L116-118