Optimizations for readers #192

acogoluegnes · 2025-07-04T08:00:12Z

A couple of optimizations for readers:

Change the signature of the sendfile callback. The callback is used in RabbitMQ to send the frame header. The callback expects now a binary result to send it along the chunk header. This saves a system call and improves performance with small chunks.
Read ahead small chunks. Instead of just reading the header and using sendfile to dispatch the chunk data, the whole chunk (and potentially the next ones) is read with the header if it can fit in a read ahead limit (approximately 4096 bytes). This saves system calls for streams with lots of small chunks and can increase dramatically the reading rate (up to 6 times for some cases). The chunk iterator also uses this read-ahead approach.

This saves a system call by sending the frame header and the chunk header at the same time. References rabbitmq/osiris#192

The callback is used in RabbitMQ to send the frame header. This commit changes the callback to expect a binary result and to send it along the chunk header. This saves a system call and improves performance with small chunks.

This saves a system call by sending the frame header and the chunk header at the same time. References rabbitmq/osiris#192

src/osiris_log.erl

This saves a system call by sending the frame header and the chunk header at the same time. References rabbitmq/osiris#192

Osiris can read ahead data in case of small chunks. This saves system calls and increases consumption rate dramatically for some streams. This is transparent for the stream protocol, but requires a small tweak for the stream queue type implementation (passing in the previous iterator when creating a new one). The read ahead is on by default but can be deactivated with to the new stream.read_ahead configuration entry (true / false). References rabbitmq/osiris#192

Osiris can read ahead data in case of small chunks. This saves system calls and increases consumption rate dramatically for some streams. This is transparent for the stream protocol, but requires a small tweak for the stream queue type implementation (passing in the previous iterator when creating a new one). The read ahead is on by default but can be deactivated with to the new stream.read_ahead configuration entry (true / false). Co-authored-by: Karl Nilsson <kjnilsson@gmail.com> References rabbitmq/osiris#192

Reading a chunk header then using the sendfile syscall to send the chunk data is not optimal for small chunks, as it amounts to 2 system calls for a small amount of bytes. This commit uses the size of the last chunk to try to read the whole chunk when the library reads the chunk header. If the last chunk data size was under 4096 bytes, the library does not only read the chunk header, but reads ahead to try to read the whole and thus to save the sendfile call later. This optimization should improve the read throughput for streams with small chunks, e.g. streams with 1-message chunks.

Start with default filter size and use actual size after non-zero value has been read. Attempt to read ahead from parse_header instead of returning the header and no data. The decision is based on the availability of previous read ahead data. Use read ahead data in chunk iterator if available

And some small fixes.

To avoid losing potential read ahead data that is filled during iteration. Comment updates.

And test read ahead switch on iterator.

by the next read-ahead.

This saves a system call by sending the frame header and the chunk header at the same time. References rabbitmq/osiris#192

Osiris can read ahead data in case of small chunks. This saves system calls and increases consumption rate dramatically for some streams. This is transparent for the stream protocol, but requires a small tweak for the stream queue type implementation (passing in the previous iterator when creating a new one). The read ahead is on by default but can be deactivated with to the new stream.read_ahead configuration entry (true / false). Co-authored-by: Karl Nilsson <kjnilsson@gmail.com> References rabbitmq/osiris#192

This saves a system call by sending the frame header and the chunk header at the same time. References rabbitmq/osiris#192 (cherry picked from commit 885e89e)

Osiris can read ahead data in case of small chunks. This saves system calls and increases consumption rate dramatically for some streams. This is transparent for the stream protocol, but requires a small tweak for the stream queue type implementation (passing in the previous iterator when creating a new one). The read ahead is on by default but can be deactivated with to the new stream.read_ahead configuration entry (true / false). Co-authored-by: Karl Nilsson <kjnilsson@gmail.com> References rabbitmq/osiris#192 (cherry picked from commit 9f162df)

This saves a system call by sending the frame header and the chunk header at the same time. References rabbitmq/osiris#192

Osiris can read ahead data in case of small chunks. This saves system calls and increases consumption rate dramatically for some streams. This is transparent for the stream protocol, but requires a small tweak for the stream queue type implementation (passing in the previous iterator when creating a new one). The read ahead is on by default but can be deactivated with to the new stream.read_ahead configuration entry (true / false). Co-authored-by: Karl Nilsson <kjnilsson@gmail.com> References rabbitmq/osiris#192

acogoluegnes added a commit to rabbitmq/rabbitmq-server that referenced this pull request Jul 4, 2025

Return stream frame header binary in dispatch chunk callback

de29b53

This saves a system call by sending the frame header and the chunk header at the same time. References rabbitmq/osiris#192

acogoluegnes mentioned this pull request Jul 4, 2025

Integrate Osiris reader optimizations for streams rabbitmq/rabbitmq-server#14189

Merged

acogoluegnes force-pushed the send-file-improvements branch 5 times, most recently from c5572f0 to bda8f17 Compare July 24, 2025 10:07

acogoluegnes changed the title ~~Update send_file callback to save a system call~~ Optimizations for readers Jul 24, 2025

kjnilsson force-pushed the send-file-improvements branch from c16bed9 to b50de25 Compare July 29, 2025 15:15

acogoluegnes added a commit to rabbitmq/rabbitmq-server that referenced this pull request Aug 1, 2025

Return stream frame header binary in dispatch chunk callback

6a8a73f

This saves a system call by sending the frame header and the chunk header at the same time. References rabbitmq/osiris#192

acogoluegnes added a commit to rabbitmq/rabbitmq-server that referenced this pull request Aug 4, 2025

Return stream frame header binary in dispatch chunk callback

a4b08fd

This saves a system call by sending the frame header and the chunk header at the same time. References rabbitmq/osiris#192

acogoluegnes added a commit to rabbitmq/rabbitmq-server that referenced this pull request Aug 4, 2025

Return stream frame header binary in dispatch chunk callback

c0aef41

This saves a system call by sending the frame header and the chunk header at the same time. References rabbitmq/osiris#192

acogoluegnes force-pushed the send-file-improvements branch from 3827fac to b561e82 Compare August 22, 2025 08:16

kjnilsson force-pushed the send-file-improvements branch 2 times, most recently from 8cce6ed to ec17a39 Compare August 26, 2025 09:29

kjnilsson pushed a commit to rabbitmq/rabbitmq-server that referenced this pull request Aug 26, 2025

Return stream frame header binary in dispatch chunk callback

34b8900

This saves a system call by sending the frame header and the chunk header at the same time. References rabbitmq/osiris#192

acogoluegnes added a commit to rabbitmq/rabbitmq-server that referenced this pull request Sep 1, 2025

Return stream frame header binary in dispatch chunk callback

6f7dcc5

This saves a system call by sending the frame header and the chunk header at the same time. References rabbitmq/osiris#192

kjnilsson force-pushed the send-file-improvements branch from ec17a39 to 182cb04 Compare September 1, 2025 14:07

acogoluegnes force-pushed the send-file-improvements branch from 208e5d7 to 48bc4ee Compare September 3, 2025 06:29

acogoluegnes added a commit to rabbitmq/rabbitmq-server that referenced this pull request Sep 4, 2025

Return stream frame header binary in dispatch chunk callback

1cdc9b9

This saves a system call by sending the frame header and the chunk header at the same time. References rabbitmq/osiris#192

Update send_file callback to save a system call

57c7b5b

The callback is used in RabbitMQ to send the frame header. This commit changes the callback to expect a binary result and to send it along the chunk header. This saves a system call and improves performance with small chunks.

acogoluegnes force-pushed the send-file-improvements branch from 5084a4f to 647e470 Compare September 4, 2025 08:49

acogoluegnes added a commit to rabbitmq/rabbitmq-server that referenced this pull request Sep 9, 2025

Return stream frame header binary in dispatch chunk callback

b79d642

This saves a system call by sending the frame header and the chunk header at the same time. References rabbitmq/osiris#192

kjnilsson requested changes Sep 19, 2025

View reviewed changes

src/osiris_log.erl Outdated Show resolved Hide resolved

acogoluegnes and others added 11 commits September 22, 2025 09:55

Refactoring

01b50a8

refactoring

25efbfd

Test read ahead

3d688da

And some small fixes.

Add function to generate a log

9e48879

revised readahead approach for chunk iterators

fef8b25

Make read-ahead an opt-out

5a0d21a

Allow read ahed buffer to be passed from iterator to log reader.

3868e51

To avoid losing potential read ahead data that is filled during iteration. Comment updates.

Add read_ahead to reader options

1b71063

And test read ahead switch on iterator.

Only enable read ahead if prior chunk can be full covered

01427d8

by the next read-ahead.

acogoluegnes force-pushed the send-file-improvements branch from bf7d7ea to 01427d8 Compare September 22, 2025 07:57

acogoluegnes marked this pull request as ready for review September 22, 2025 08:51

acogoluegnes merged commit 808f1f2 into main Sep 22, 2025
4 checks passed

acogoluegnes deleted the send-file-improvements branch September 22, 2025 09:01

mergify bot mentioned this pull request Sep 22, 2025

Integrate Osiris reader optimizations for streams (backport #14189) rabbitmq/rabbitmq-server#14586

Merged

the-mikedavis mentioned this pull request Oct 14, 2025

Add abstractions for tiered storage of stream data #196

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Optimizations for readers #192

Optimizations for readers #192

Uh oh!

acogoluegnes commented Jul 4, 2025 •

edited

Loading

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Optimizations for readers #192

Optimizations for readers #192

Uh oh!

Conversation

acogoluegnes commented Jul 4, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

acogoluegnes commented Jul 4, 2025 •

edited

Loading