Skip to content

fix(s3s-fs): set TCP_NODELAY on every accepted connection - #3

Closed
pmoust wants to merge 1 commit into
mainfrom
fix/s3s-fs-tcp-nodelay
Closed

pmoust wants to merge 1 commit into
mainfrom
fix/s3s-fs-tcp-nodelay

Conversation

@pmoust

@pmoust pmoust commented Nov 2, 2026

Copy link
Copy Markdown
Member

Validation only; upstream port of the ravel.2 TCP_NODELAY commit.

The accept loop served each socket with Nagle on. A response that hyper
writes as a head and then a body holds the body until the client ACKs
the head, and a Linux client delays that ACK by 40 ms. Every request
whose response went out in two writes paid it.

Measured on the Ravel MetricsBench ci stack (NOFireAI/ravel issue 3561):
a warm selector query took 50.3 to 50.9 ms against s3s-fs, and 9.3 to
9.6 ms with quickack set in the client's network namespace. A tcpdump
of port 9000 showed one 40.4 to 41.0 ms gap per query, each after a
204-byte segment from s3s-fs and ended by the client's ACK.

A socket that refuses the option is served without it.

Signed-off-by: Panagiotis Moustafellos <pmoust@nofire.ai>
@pmoust

pmoust commented Nov 2, 2026

Copy link
Copy Markdown
Member Author

Validation done; proposed upstream as s3s-project#1017

@pmoust pmoust closed this Nov 2, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant