[
https://issues.apache.org/jira/browse/THRIFT-6159?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Jens Geyer reassigned THRIFT-6159:
----------------------------------
Assignee: Léo - Tinkeam
> Rust: Use faster varint library - #3739
> ---------------------------------------
>
> Key: THRIFT-6159
> URL: https://issues.apache.org/jira/browse/THRIFT-6159
> Project: Thrift
> Issue Type: Improvement
> Components: Rust - Library
> Affects Versions: 0.24.0
> Reporter: Léo - Tinkeam
> Assignee: Léo - Tinkeam
> Priority: Minor
> Original Estimate: 168h
> Remaining Estimate: 168h
>
> Changed the Rust varint library from integer-encoding to varint-rs, [my
> tests|https://github.com/Leo-Tinkeam/bench-rust-varint] on differents library
> reveal that varint-rs can be up to 2 times faster.
> Also tested thrift before and after
> [here|https://github.com/Leo-Tinkeam/bench-thrift], 20% improvements in
> release mode (`cargo run --release`) but 60% slower with dev (`cargo run`).
> This may be an issue but I think that dev time don't matters that much.
> The generated output.bin are exactly the same before and after (tested with
> `cmp master.bin before.bin`).
> My usage is not the basic usage of the library, integer-encoding is using
> `write_all` once on `transport` but varint-rs is doing it for each byte,
> which is slower with our implementation of `transport.write_all`, I am
> writing to a buffer with varint-rs in order to use `write_all` only once at
> the end.
> I think that transport should use a fixed size buffer (something like 1ko ?)
> and the buffer that is up to 10o that I implemented here will became
> irrelevant. I'm not sure about this last observation but I may investigate if
> you want me to.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
