ε-serde

ε-serde is a Rust framework for ε-copy serialization and deserialization.

Why

Large immutable data structures need time to be deserialized using the serde approach. A possible solution for this problem is given by frameworks such as Abomonation, rkiv, and zerovec, which provide zero-copy deserialization: the stream of bytes serializing the data structure can be used directly as a Rust structure. In particular, this approach makes it possible to map into memory an on-disk data structure, making it available instantly. It also makes it possible to load the data structure in a memory region with particular attributes, such as transparent huge pages on Linux. Even when using standard memory load and deserialization happen much faster as the entire structure can be loaded with a single read operation.

ε-serde has the same goals as the zero-copy frameworks above but provides different tradeoffs.

How

Since in these data structures typically most of the data is given by large chunks of memory in the form of slices or vectors, at deserialization time one can build quickly a proper Rust structure whose referenced memory, however, is not copied. We call this approach ε-copy deserialization, as typically a minuscule fraction of the serialized data is copied to build the structure. The result is similar to that of the frameworks above, but the performance of the deserialized structure will be identical to that of a standard, in-memory Rust structure, as references are resolved at deserialization time.

We provide procedural macros implementing serialization and deserialization methods, basic (de)serialization for primitive types, vectors, etc., convenience memory-mapping methods based on mmap_rs, and a MemCase structure that couples a deserialized structure with its backend (e.g., a slice of memory or a memory-mapped region).

Who

Tommaso Fontana, while working at INRIA under the supervision of Stefano Zacchiroli, came up with the basic idea for ε-serde, that is, replacing structures with equivalent references. The code was developed jointly with Sebastiano Vigna, who came up with the MemCase and the ZeroCopy/DeepCopy logic.

Cons

These are the main limitations you should be aware of before choosing to use ε-serde:

• Your types cannot contain references. For example, you cannot use ε-serde on a tree.

• While we provide procedural macros that implement serialization and deserialization, they require that your type is written and used in a specific way; in particular, the fields you want to ε-copy must be type parameters implementing DeserializeInner, to which a deserialized type is associated. For example, we provide implementations for Vec<T>/Box<[T]>, where T is zero-copy, or String/Box<str>, which have associated deserialized type &[T] or &str, respectively. Vectors and boxed slices of types that are not zero-copy will be deserialized recursively in memory instead.

• After deserialization of a type T, you will obtain an associated deserialized type DeserType<'_,T>, which will usually reference the underlying serialized support (e.g., a memory-mapped region); hence the need for a lifetime. If you need to store the deserialized structure in a field of a new structure you will need to couple permanently the deserialized structure with its serialized support, which is obtained by putting it in a MemCase using the convenience methods Deserialize::load_mem, Deserialize::load_mmap, and Deserialize::mmap. A MemCase will deref to its contained type, so it can be used transparently as long as fields and methods are concerned, but if your original type is T the field of the new structure will have to be of type MemCase<DeserType<'static, T>>, not T.

Pros

• Almost instant deserialization with minimal allocation, provided that you designed your type following the ε-serde guidelines or that you use standard types.

• The structure you get by deserialization is the same structure you serialized, except that type parameters will be replaced by their associated deserialization type (e.g., vectors will become become references to slices). This is not the case with rkiv, which requires you to reimplement all methods on the deserialized type.

• The structure you get by deserialization has exactly the same performance as the structure you serialized. This is not the case with zerovec or rkiv.

• You can deserialize from read-only supports, as all dynamic information generated at deserialization time is stored in newly allocated memory. This is not the case with Abomonation.

Example: Zero copy of standard types

Let us start with the simplest case: data that can be zero-copy deserialized. In this case, we serialize an array of a thousand zeros, and get back a reference to such an array:

# fn main() -> Result<(), Box<dyn std::error::Error>> {
use epserde::prelude::*;

let s = [0_usize; 1000];

// Serialize it
let mut file = std::env::temp_dir();
file.push("serialized0");
s.serialize(&mut std::fs::File::create(&file)?)?;
// Load the serialized form in a buffer
let b = std::fs::read(&file)?;

// The type of t will be inferred--it is shown here only for clarity
let t: &[usize; 1000] =
    <[usize; 1000]>::deserialize_eps(b.as_ref())?;

assert_eq!(s, *t);

// You can derive the deserialization type, with a lifetime depending on b
let t: DeserType<'_, [usize; 1000]> =
    <[usize; 1000]>::deserialize_eps(b.as_ref())?;

assert_eq!(s, *t);

// This is a traditional deserialization instead
let t: [usize; 1000] = 
    <[usize; 1000]>::deserialize_full(
        &mut std::fs::File::open(&file)?
    )?;
assert_eq!(s, t);

// In this case we map the data structure into memory
let u: MemCase<&[usize; 1000]> = 
    <[usize; 1000]>::mmap(&file, Flags::empty())?;

assert_eq!(s, **u);

// When using a MemCase, the lifetime of the derived deserialization type is 'static
let u: MemCase<DeserType<'static, [usize; 1000]>> = 
    <[usize; 1000]>::mmap(&file, Flags::empty())?;

assert_eq!(s, **u);
#     Ok(())
# }

Note how we serialize an array, but we deserialize a reference. The reference points inside b, so there is no copy performed. The call to deserialize_full creates a new array instead. The third call maps the data structure into memory and returns a MemCase that can be used transparently as a reference to the array; moreover, the MemCase can be passed to other functions or stored in a structure field, as it contains both the structure and the memory-mapped region that supports it.

The type alias DeserType can be used to derive the deserialized type associated with a type. It contains a lifetime, which is the lifetime of the memory region containing the serialized data. When deserializing into a MemCase, however, the lifetime is 'static, as MemCase is an owned type.

Examples: ε-copy of standard structures

Zero-copy deserialization is not that interesting because it can be applied only to data whose memory layout and size are fixed and known at compile time. This time, let us serialize a Vec containing a thousand zeros: ε-serde will deserialize its associated deserialization type, which is a reference to a slice.

# fn main() -> Result<(), Box<dyn std::error::Error>> {
use epserde::prelude::*;

let s = vec![0; 1000];

// Serialize it
let mut file = std::env::temp_dir();
file.push("serialized1");
s.serialize(&mut std::fs::File::create(&file)?)?;
// Load the serialized form in a buffer
let b = std::fs::read(&file)?;

// The type of t will be inferred--it is shown here only for clarity
let t: DeserType<'_, Vec<usize>> =
    <Vec<usize>>::deserialize_eps(b.as_ref())?;

assert_eq!(s, *t);

// This is a traditional deserialization instead
let t: Vec<usize> = 
    <Vec<usize>>::load_full(&file)?;
assert_eq!(s, t);

// In this case we map the data structure into memory
let u: MemCase<DeserType<'static, Vec<usize>>> = 
    <Vec<usize>>::mmap(&file, Flags::empty())?;
assert_eq!(s, **u);
#     Ok(())
# }

Note how we serialize a vector, but we deserialize a reference to a slice; the same would happen when serializing a boxed slice. The reference points inside b, so there is very little copy performed (in fact, just a field containing the length of the slice). All this is because usize is a zero-copy type. Note also that we use the convenience method Deserialize::load_full.

If your code must work both with the original and the deserialized version, however, it must be written for a trait that is implemented by both types, such as AsRef<[usize]>.

Example: Zero-copy structures

You can define your types to be zero-copy, in which case they will work like usize in the previous examples. This requires the structure to be made of zero-copy fields, and to be annotated with #[zero_copy] and #[repr(C)] (which means that you will lose the possibility that the compiler reorders the fields to optimize memory usage):

# fn main() -> Result<(), Box<dyn std::error::Error>> {
use epserde::prelude::*;
use epserde_derive::*;

#[derive(Epserde, Debug, PartialEq, Copy, Clone)]
#[repr(C)]
#[zero_copy]
struct Data {
    foo: usize,
    bar: usize,
}

let s = vec![Data { foo: 0, bar: 0 }; 1000];

// Serialize it
let mut file = std::env::temp_dir();
file.push("serialized2");
s.serialize(&mut std::fs::File::create(&file)?)?;
// Load the serialized form in a buffer
let b = std::fs::read(&file)?;

// The type of t will be inferred--it is shown here only for clarity
let t: DeserType<'_, Vec<Data>> =
    <Vec<Data>>::deserialize_eps(b.as_ref())?;

assert_eq!(s, *t);

// This is a traditional deserialization instead
let t: Vec<Data> = 
    <Vec<Data>>::load_full(&file)?;
assert_eq!(s, t);

// In this case we map the data structure into memory
let u: MemCase<DeserType<'static, Vec<Data>>> = 
    <Vec<Data>>::mmap(&file, Flags::empty())?;
assert_eq!(s, **u);
#     Ok(())
# }

If a structure is not zero-copy, vectors of structures will be always deserialized into vectors.

Example: Structures with parameters

More flexibility can be obtained by defining structures with fields whose types are defined by parameters. In this case, ε-serde will deserialize the structure replacing its type parameters with the associated deserialized type.

Let us design a structure that will contain an integer, which will be copied, and a vector of integers that we want to ε-copy:

# fn main() -> Result<(), Box<dyn std::error::Error>> {
use epserde::prelude::*;
use epserde_derive::*;

#[derive(Epserde, Debug, PartialEq)]
struct MyStruct<A> {
    id: isize,
    data: A,
}

// Create a structure where A is a Vec<isize>
let s: MyStruct<Vec<isize>> = MyStruct { id: 0, data: vec![0, 1, 2, 3] };
// Serialize it
let mut file = std::env::temp_dir();
file.push("serialized3");
s.store(&file);
// Load the serialized form in a buffer
let b = std::fs::read(&file)?;

// The type of t will be inferred--it is shown here only for clarity
let t: DeserType<'_, MyStruct<Vec<isize>>> = 
    <MyStruct<Vec<isize>>>::deserialize_eps(b.as_ref())?;

assert_eq!(s.id, t.id);
assert_eq!(s.data, Vec::from(t.data));

// This is a traditional deserialization instead
let t: MyStruct<Vec<isize>> = 
    <MyStruct::<Vec<isize>>>::load_full(&file)?;
assert_eq!(s, t);

// In this case we map the data structure into memory
let u: MemCase<DeserType<'static, MyStruct<Vec<isize>>>> = 
    <MyStruct::<Vec<isize>>>::mmap(&file, Flags::empty())?;
assert_eq!(s.id, u.id);
assert_eq!(s.data, u.data.as_ref());
#     Ok(())
# }

Note how the field originally containing a Vec<isize> now contains a &[isize] (this replacement is generated automatically). The reference points inside b, so there is no need to copy the field. Nonetheless, deserialization creates a new structure MyStruct, ε-copying the original data. The second call creates a full copy instead. We can write methods for our structure that will work for the ε-copied version: we just have to take care that they are defined in a way that will work both on the original type parameter and on its associated deserialized type; we can also use type to reduce the clutter:

# fn main() -> Result<(), Box<dyn std::error::Error>> {
use epserde::prelude::*;
use epserde_derive::*;

#[derive(Epserde, Debug, PartialEq)]
struct MyStructParam<A> {
    id: isize,
    data: A,
}

/// This method can be called on both the original and the ε-copied structure
impl<A: AsRef<[isize]>> MyStructParam<A> {
    fn sum(&self) -> isize {
        self.data.as_ref().iter().sum()
    }
}

type MyStruct = MyStructParam<Vec<isize>>;

// Create a structure where A is a Vec<isize>
let s = MyStruct { id: 0, data: vec![0, 1, 2, 3] };
// Serialize it
let mut file = std::env::temp_dir();
file.push("serialized4");
s.store(&file);
// Load the serialized form in a buffer
let b = std::fs::read(&file)?;
let t = MyStruct::deserialize_eps(b.as_ref())?;
// We can call the method on both structures
assert_eq!(s.sum(), t.sum());

let t = <MyStruct>::mmap(&file, Flags::empty())?;

// t works transparently as a MyStructParam<&[isize]>
assert_eq!(s.id, t.id);
assert_eq!(s.data, t.data.as_ref());
assert_eq!(s.sum(), t.sum());
#     Ok(())
# }

Example: Deep-copy structures with internal parameters

Internal parameters, that is, parameters used by the types of your fields but that do not represent the type of your fields are left untouched. However, to be serializable they must be classified as deep-copy or zero-copy, and in the first case they must have a 'static lifetime. For example,

# fn main() -> Result<(), Box<dyn std::error::Error>> {
use epserde::prelude::*;
use epserde_derive::*;

#[derive(Epserde, Debug, PartialEq)]
struct MyStruct<A: DeepCopy + 'static>(Vec<A>);

// Create a structure where A is a Vec<isize>
let s: MyStruct<Vec<isize>> = MyStruct(vec![vec![0, 1, 2, 3]]);
// Serialize it
let mut file = std::env::temp_dir();
file.push("serialized4");
s.store(&file);
// Load the serialized form in a buffer
let b = std::fs::read(&file)?;

// The type of t is unchanged
let t: MyStruct<Vec<isize>> = 
    <MyStruct<Vec<isize>>>::deserialize_eps(b.as_ref())?;
#     Ok(())
# }

Note how the field originally of type Vec<Vec<isize>> remains of the same type.

Example: Zero-copy structures with parameters

For zero-copy structure, things are slightly different because types are not substituted, even if they represent the type of your fields. So all parameters must be zero-copy. For example,

# fn main() -> Result<(), Box<dyn std::error::Error>> {
use epserde::prelude::*;
use epserde_derive::*;

#[derive(Epserde, Debug, PartialEq, Clone, Copy)]
#[repr(C)]
#[zero_copy]
struct MyStruct<A: ZeroCopy> {
    data: A,
}

// Create a structure where A is a Vec<isize>
let s: MyStruct<i32> = MyStruct { data: 0 };
// Serialize it
let mut file = std::env::temp_dir();
file.push("serialized5");
s.store(&file);
// Load the serialized form in a buffer
let b = std::fs::read(&file)?;

// The type of t is unchanged
let t: &MyStruct<i32> = 
    <MyStruct<i32>>::deserialize_eps(b.as_ref())?;
#     Ok(())
# }

Note how the field originally of type i32 remains of the same type.

Example: Enums

Enums are supported, but there are two caveats: first, if you want them to be zero-copy, they must be repr(C), and thus you will lose the possibility that the compiler optimize their memory representation; second, if you have type parameters that are not used by all variants you must be careful to specify always the same type parameter when serializing and deserializing. This is obvious for non-enum types, but with enum types with default type parameters it can become tricky. For example,

# fn main() -> Result<(), Box<dyn std::error::Error>> {
use epserde::prelude::*;
use epserde_derive::*;

#[derive(Epserde, Debug, PartialEq, Clone, Copy)]
enum Enum<T=Vec<usize>> {
    A,
    B(T),
}

// This enum has T=Vec<i32> by type inference
let e = Enum::B(vec![0, 1, 2, 3]);
// Serialize it
let mut file = std::env::temp_dir();
file.push("serialized6");
e.store(&file);
// Deserializing using just Enum will fail, as the type parameter 
// by default is Vec<usize>
assert!(<Enum>::load_full(&file).is_err());
#     Ok(())
# }

Example: sux-rs

The sux-rs crate provides several data structures that use ε-serde.

Design

Every type serializable with ε-serde has two features that are in principle orthogonal, but that in practice often condition one another:

• the type has an associated deserialization type, which is the type you obtain upon deserialization;
• the type can be either ZeroCopy or DeepCopy; it can also be neither.

There is no constraint on the associated deserialization type: it can be literally anything. In general, however, one tries to have a deserialization type that is somewhat compatible with the original type, in the sense that they both satisfy a trait for which implementations can be written: for example, ε-serde deserializes vectors as references to slices, so implementations can be written for references to slices and will work both on the original and the deserialized type. And, in general, ZeroCopy types deserialize into themselves.

Being ZeroCopy or DeepCopy decides instead how the type will be treated when serializing and deserializing sequences, such as arrays, slices, boxed slices, and vectors. Sequences of zero-copy types are deserialized using a reference, whereas sequences of deep-copy types are recursively deserialized in allocated memory (to sequences of the associated deserialization types). It is important to remark that you cannot serialize a sequence whose elements are of a type that is neither ZeroCopy nor DeepCopy (see the CopyType documentation for a deeper explanation).

Logically, zero-copy types should be deserialized to references, and this indeed happens in most cases, and certainly in the derived code: however, primitive types are always fully deserialized. There are two reasons behind this non-orthogonal choice:

• primitive types occupy so little space that deserializing them as a reference is not efficient;
• if a type parameter T is a primitive type, writing generic code for AsRef<T> is really not nice;
• deserializing primitive types to a reference would require further padding to align them.

Since this is true only of primitive types, when deserializing a 1-tuple containing a primitive type one obtains a reference (and indeed this workaround can be used if you really need to deserialize a primitive type as a reference). The same happens if you deserialize a zero-copy struct containing a single field of primitive type.

Deep-copy types instead are serialized and deserialized recursively, field by field. The basic idea in ε-serde is that if a field has a type that is a parameter, during ε-copy deserialization the type will be replaced with its deserialization type. Since the deserialization type is defined recursively, replacement can happen at any depth level. For example, a field of type A = Vec<Vec<Vec<usize>>> will be deserialized as a A = Vec<Vec<&[usize]>>.

This approach makes it possible to write ε-serde-aware structures that hide from the user the substitution. A good example is the BitFieldVec structure from sux-rs, which exposes an array of fields of fixed bit width using (usually) a Vec<usize> as a backend; except for extension methods, all methods of BitFieldVec come from the trait BitFieldSlice. If you have your own struct and one of the fields is of type A, when serializing your struct with A equal to BitFieldVec<Vec<usize>>, upon ε-copy deserialization you will get a version of your struct with BitFieldVec<&[usize]>. All this will happen under the hood because BitFieldVec is ε-serde-aware, and in fact you will not even notice the difference if you access both versions using the trait BitFieldSlice.

MemDbg / MemSize

All ε-serde structures implement the MemDbg and MemSize traits.

Derived and hand-made implementations

We strongly suggest using the procedural macro Epserde to make your own types serializable and deserializable. Just invoking the macro on your structure will make it fully functional with ε-serde. The attribute #[zero_copy] can be used to make a structure zero-copy, albeit it must satisfy a few prerequisites.

You can also implement manually the traits CopyType, MaxSizeOf, TypeHash, ReprHash, SerializeInner, and DeserializeInner, but the process is error-prone, and you must be fully aware of ε-serde's conventions. The procedural macro TypeInfo can be used to generate automatically at least MaxSizeOf, TypeHash, and ReprHash automatically.

Acknowledgments

This software has been partially supported by project SERICS (PE00000014) under the NRRP MUR program funded by the EU - NGEU, and by project ANR COREGRAPHIE, grant ANR-20-CE23-0002 of the French Agence Nationale de la Recherche.