rust: Use minimum internal buffer size of 128b

[silverjam]

Benchmarks, jemalloc without 128b minimum buffer size:

Benchmark #1: sbp2json <test_data/benchmark.sbp | json2sbp >/dev/null
  Time (mean ± σ):      1.755 s ±  0.076 s    [User: 2.689 s, System: 0.532 s]
  Range (min … max):    1.585 s …  1.922 s    50 runs

Using jemalloc with 128b minimum internal buffer size:

Benchmark #1: sbp2json <test_data/benchmark.sbp | json2sbp >/dev/null
  Time (mean ± σ):     718.1 ms ±  33.9 ms    [User: 1.085 s, System: 0.170 s]
  Range (min … max):   658.3 ms … 839.5 ms    50 runs

Using mimalloc with 128b minimum internal buffer size:

Benchmark #1: sbp2json <test_data/benchmark.sbp | json2sbp >/dev/null
  Time (mean ± σ):     642.9 ms ±  33.9 ms    [User: 989.1 ms, System: 160.9 ms]
  Range (min … max):   603.4 ms … 785.0 ms    50 runs

MiMalloc also has the benefit that it works for more than just Linux.

[notoriaga]

Would something like this work?

let mut v = Vec::with_capacity(BUFLEN / mem::size_of::<T>());

it should over-allocate less often. but I think the buffer sizes would vary more

[silverjam]

I tried this which didn't seem to have any impact:

diff --git a/rust/sbp/src/wire_format.rs b/rust/sbp/src/wire_format.rs
index afdd09fd..3d3639d3 100644
--- a/rust/sbp/src/wire_format.rs
+++ b/rust/sbp/src/wire_format.rs
@@ -53,7 +53,7 @@ where
     }

     fn parse_unchecked<B: Buf>(buf: &mut B) -> Self {
-        let mut v = Vec::with_capacity(BUFLEN);
+        let mut v = Vec::with_capacity(64*(1+(mem::size_of::<T>() / 64)));
         while buf.remaining() >= T::MIN_LEN {
             v.push(T::parse_unchecked(buf));
         }

Benchmark:

Benchmark #1: sbp2json <test_data/benchmark.sbp | json2sbp >/dev/null
  Time (mean ± σ):     610.7 ms ±  18.9 ms    [User: 936.4 ms, System: 161.0 ms]
  Range (min … max):   561.0 ms … 684.1 ms    50 runs

[silverjam]

Smaller internal buffer size of 64 bytes was a bit slower:

Benchmark #1: sbp2json <test_data/benchmark.sbp | json2sbp >/dev/null
  Time (mean ± σ):     629.7 ms ±  18.2 ms    [User: 979.7 ms, System: 151.9 ms]
  Range (min … max):   584.6 ms … 678.3 ms    50 runs

[silverjam]

Buffer size of 256:

Benchmark #1: sbp2json <test_data/benchmark.sbp | json2sbp >/dev/null
  Time (mean ± σ):     624.5 ms ±  18.7 ms    [User: 962.8 ms, System: 159.4 ms]
  Range (min … max):   587.7 ms … 698.3 ms    50 runs

[notoriaga]

Hmm interesting, I guess T is typically small anyway

[notoriaga]

Any reason to not have a version?

[silverjam]

No version is what the docs for the package recommended but they didn't give a rational, I can pin it

[notoriaga]

Oops was on the wrong branch for that push, I don't think it actually pushed anything though commit changed but I don't see what actually did

[silverjam]

Oops was on the wrong branch for that push, I don't think it actually pushed anything though commit changed but I don't see what actually did

no worries, I figured it out