zerovec slower than cow?

[Thomblin]

I found this project through https://manishearth.github.io/blog/2022/08/03/zero-copy-2-zero-copy-all-the-things/ and was trying some benchmark using quick_xml::de::from_str expecting that zerotovec is much faster then using String or Cow regarding the numbers on https://docs.rs/zerovec/latest/zerovec/

I am just wondering: Did I misunderstand something or did I implement something the wrong way?

Basically what I am doing is:

let r: Root = quick_xml::de::from_str(xml).unwrap();

where Root is a serde struct like

#[derive(Serialize, Deserialize)]
pub struct Root{
    #[serde(rename = "$text")]
    pub text: Option<String>,
    #[serde(rename = "@Status")]
    pub status: String,
}

for a 5 to 80 MB xml file containing around 60 different nested elements.

For the benchmark I replace all String occurences with either
Cow<'a, str> or VarZeroVec<'data, str>
and adding the serde(borrow) directive

However I am suprised by the benchmark:

test benches::bench_parse_to_cow ... bench: 368,994 ns/iter (+/- 175,561)
test benches::bench_parse_to_string ... bench: 400,821 ns/iter (+/- 174,372)
test benches::bench_parse_to_zerovec ... bench: 620,131 ns/iter (+/- 208,128)

So my question is: Is this expected, or do I need to change something to gain the full advantage of zerotovec?

[Manishearth]

I don't understand, VarZeroVec<str> is not the equivalent of Cow<str> and String. It's the equivalent of Vec<String> (which cannot be represented in Cow in a way that will not allocate on deserialization).

There's no advantage to using ZeroVec over Cow: ZeroVec supports doing more things in zero-copy. If you can do it in Cow without allocating you don't need this crate.

[sffc]

XML supports character escape sequences, so it is not generally possible to get a &str, and Serde does not support conditionally parsing as borrowed or owned based on input (citation), so it won't be possible to improve deserialization performance of XML using zerovec.

[Manishearth]

Oh, I didn't even notice that.

Yeah zerovec only helps in situations where zero copy deserialization is possible, which typically means binary serialization formats.

[Thomblin]

Thanks a lot for your input. Still learning how the details of Rust are working. I guess I need to read some more

@Manishearth I am not sure if I understand your last comment correctly. Does quick_xml::de with serde use zero copy deserialization if I choose Cow<'a, str> as type?

[Manishearth]

I believe it does not.