[Task][Go SDK][Perf]: Make and use a general []byte buffer allocation pool for the SDK.

[lostluck]

What needs to happen?

Replace all small []byte, or bytes.Buffer uses with a central sync.Pool allocated "buffer" package, like

https://cs.opensource.google/go/x/exp/+/f062dba9:slog/internal/buffer/buffer.go;bpv=1;bpt=1

or

https://github.com/golang/go/blob/master/src/log/log.go#L165

https://pkg.go.dev/sync#Pool are weak reference pools meaning that buffers in the pool may be freed at during the next GC if appropriate.

While some unsafe strategies were used as well, through the ioutilx package read and write calls to avoid allocations, this remains slightly risky, so uses there are a reasonable priority.

But generally, uses of bytes.Buffer or arbitrary lengths but known lifetime uses of make([]byte)

Some bytes.Buffer replacements would require this new Buffer to support the io.Reader interface as well.

In particular, targeting the uses in the exec and harness packages would likely be the most fruitful.

It would be worth having a moderate scale benchmark to validate the before and after performance of this change, with appropriate heap and CPU profiles.

---

Notionally the benefits would be a reduction in byte buffer allocations to the heap through re-use, while also enabling certain levels of inlining, though this would require more explicit use of the new buffer type instead of using the interfaces.

Note that Go also remains a garbage collected language, so not all uses from the pool need to be freed. They can leak and be GC'd normally.

Issue Priority

Priority: 3 (nice-to-have improvement)

Issue Components

• Component: Python SDK
• Component: Java SDK
• Component: Go SDK
• Component: Typescript SDK
• Component: IO connector
• Component: Beam examples
• Component: Beam playground
• Component: Beam katas
• Component: Website
• Component: Spark Runner
• Component: Flink Runner
• Component: Samza Runner
• Component: Twister2 Runner
• Component: Hazelcast Jet Runner
• Component: Google Cloud Dataflow Runner

[yyy1000]

I want to have a try. :)

[lostluck]

Great!
@yyy1000 see beam.apache.org/contribute for instructions on using the bot infra to get this assigned to you.

FYI, this only needs to be done for Writes. There's no value in changing how reads are done, since bytes.Reader already exists.

Benchmarks are required for anything being changed up.

[yyy1000]

.take-issue

[lostluck]

Oh! And just a heads up, I wouldn't change anything about the harness/datamgr.go code in the near term (next 2-3 weeks). I'm working on fixing that layer so we can support Timers, and that requires changing a lot of the handling code around the reads and the writes.

(Though that code has been optimmized to re-use its buffer already: https://issues.apache.org/jira/browse/BEAM-13082 , which will be fine for batch, if not for streaming...)

[yyy1000]

I came across a solution to this, but I don't know whether my understanding is accurate. Could u please help me to review it and give some comments and suggestions? Thanks!

1. Create a package named "buffer". This package will contain the implementation of the allocated buffer by sync.Pool.
2. Define a struct named Buffer in the "buffer" package that wraps a []byte slice.
3. Implement a "GetBuffer" function in the "buffer" package that returns a new instance of the Buffer struct from the sync.Pool, and a "PutBuffer" function in the "buffer" package that returns a used Buffer struct back to the sync.Pool.
4. Replace all small []byte or bytes.Buffer uses in go sdk with calls to "GetBuffer" and "PutBuffer".

[lostluck]

Please look at the linked code from the slog package from the start of the issue. There's no need to wrap in a struct, and the type can self return to the pool via a method call.

…

On Tue, Feb 28, 2023, 6:47 AM Junhao Liu ***@***.***> wrote: I came across a solution to this, but I don't know whether my understanding is accurate. Could u please help me to review it and give some comments and suggestions? Thanks! 1. Create a package named "buffer". This package will contain the implementation of the allocated buffer by sync.Pool. 2. Define a struct named Buffer in the "buffer" package that wraps a []byte slice. 3. Implement a "GetBuffer" function in the "buffer" package that returns a new instance of the Buffer struct from the sync.Pool, and a "PutBuffer" function in the "buffer" package that returns a used Buffer struct back to the sync.Pool. 4. Replace all small []byte or bytes.Buffer uses in go sdk with calls to "GetBuffer" and "PutBuffer". — Reply to this email directly, view it on GitHub <#25522 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADKDOFIJRU7CAHEGHVWAQVLWZYFYBANCNFSM6AAAAAAU64XZNE> . You are receiving this because you authored the thread.Message ID: ***@***.***>