zommiommy / mem_dbg-rs

A trait + derive to recursively compute / print the memory usage of a data structure.

mem_dbg

Traits and associated procedural macros to display recursively the layout and memory usage of a value.

The trait MemDbg can be used to display the recursive layout of a value, together with the size of each part and the associated padding bytes. We provide implementations for most basic types, a derive macro for structs and enums whose fields implement MemDbg, and support for a few other crates.

To compute the size, we provide the trait MemSize and a derive macro that can be used to compute the size of a value in bytes as the standard library function std::mem::size_of returns the stack size of a type in bytes, but it does not take into consideration heap memory.

Why MemSize

Other traits partially provide the functionality of MemSize, but either they require implementing manually a trait, which is prone to error, or they do not provide the flexibility necessary for MemDbg. Most importantly, MemSize uses the type system to avoid iterating over the content of a container (a vector, etc.) when it is not necessary, making it possible to compute instantly the size of values occupying hundreds of gigabytes of heap memory.

This is the result of the benchmark bench_hash_map contained in the examples directory. It builds a hash map with a hundred million entries and then measure its heap size:

Allocated:    2281701509
get_size:     1879048240 152477833 ns
deep_size_of: 1879048240 152482000 ns
size_of:      2281701432 152261958 ns
mem_size:     2281701424 209 ns

The first line is the number of bytes allocated by the program as returned by cap. Then, we display the result of get-size, deepsize, size-of, and our own MemSize. Note that the first two crates are just measuring the space used by the items, and not by the data structure (i.e., they are not taking into account the load factor and the power-of-two size constraint of the hash map). Moreover, all other crates are about six orders of magnitude slower than our implementation, due to the necessity to iterate over all elements.

Padding

The trait MemDbg is useful to display the layout of a value and understand how much memory is used by each part. In particular, it exploits the new stable macro [std::mem::offset_of] to display the padding of each field in square brackets; moreover, the flag DbgFlags::RUST_LAYOUT makes it possible to display structures in the layout used by the Rust compiler, rather than that given by declaration order.

These features are also available for enums using the feature offset_of_enum, which however needs the nightly compiler, as it enables the unstable features offset_of_enum and offset_of_nested.

Features

• offset_of_enum: support for padding and for the DbgFlags::RUST_LAYOUT flag for enums. Requires the nightly compiler as it enables the unstable features offset_of_enum and offset_of_nested. Calling mem_dbg with the flag DbgFlags::RUST_LAYOUT without this feature enabled will result in a panic.
• half: support for the half crate.
• maligned: support for the maligned crate.
• mmap-rs: support for the mmap-rs crate.
• rand: support for the rand crate.

Example

# #![cfg_attr(feature = "offset_of_enum", feature(offset_of_enum, offset_of_nested))]
# fn main() -> Result<(), Box<dyn std::error::Error>> {
use mem_dbg::*;

#[derive(MemSize, MemDbg)]
struct Struct<A, B> {
    a: A,
    b: B,
    test: isize,
}

#[derive(MemSize, MemDbg)]
struct Data<A> {
    a: A,
    b: Vec<i32>,
    c: (u8, String),
}

#[derive(MemSize, MemDbg)]
enum TestEnum {
    Unit,
    Unit2(),
    Unit3 {},
    Unnamed(usize, u8),
    Named { first: usize, second: u8 },
}

let b = Vec::with_capacity(100);

let s = Struct {
    a: TestEnum::Unnamed(0, 16),
    b: Data {
        a: vec![0x42_u8; 700],
        b,
        c: (1, "foo".to_owned()),
    },
    test: -0xbadf00d,
};

println!("size:     {}", s.mem_size(SizeFlags::default()));
println!("capacity: {}", s.mem_size(SizeFlags::CAPACITY));
println!();

s.mem_dbg(DbgFlags::empty())?;

println!();

println!("size:     {}", s.mem_size(SizeFlags::default()));
println!("capacity: {}", s.mem_size(SizeFlags::CAPACITY));
println!();

s.mem_dbg(DbgFlags::default() | DbgFlags::CAPACITY | DbgFlags::HUMANIZE)?;

#[cfg(feature = "offset_of_enum")]
{
    println!();

    println!("size:     {}", s.mem_size(SizeFlags::default()));
    println!("capacity: {}", s.mem_size(SizeFlags::CAPACITY));
    println!();

    s.mem_dbg(DbgFlags::empty() | DbgFlags::RUST_LAYOUT)?;
}
# Ok(())
# }

The previous program prints:

size:     807
capacity: 1207

807 B ⏺
 16 B ├╴a
      │ ├╴Variant: Unnamed
  8 B │ ├╴0
  1 B │ ╰╴1
783 B ├╴b
724 B │ ├╴a
 24 B │ ├╴b
 35 B │ ╰╴c
  1 B │   ├╴0 [7B]
 27 B │   ╰╴1
  8 B ╰╴test

size:     807
capacity: 1207

1.207 kB 100.00% ⏺: readme::main::Struct<readme::main::TestEnum, readme::main::Data<alloc::vec::Vec<u8>>>
   16  B   1.33% ├╴a: readme::main::TestEnum
                 │ ├╴Variant: Unnamed
    8  B   0.66% │ ├╴0: usize
    1  B   0.08% │ ╰╴1: u8
1.183 kB  98.01% ├╴b: readme::main::Data<alloc::vec::Vec<u8>>
  724  B  59.98% │ ├╴a: alloc::vec::Vec<u8>
  424  B  35.13% │ ├╴b: alloc::vec::Vec<i32>
   35  B   2.90% │ ╰╴c: (u8, alloc::string::String)
    1  B   0.08% │   ├╴0: u8 [7B]
   27  B   2.24% │   ╰╴1: alloc::string::String
    8  B   0.66% ╰╴test: isize

If run with the feature offset_of_enum, it prints:

size:     807
capacity: 1207

807 B ⏺
 16 B ├╴a
      │ ├╴Variant: Unnamed
  8 B │ ├╴0
  1 B │ ╰╴1 [6B]
783 B ├╴b
724 B │ ├╴a
 24 B │ ├╴b
 35 B │ ╰╴c
  1 B │   ├╴0 [7B]
 27 B │   ╰╴1
  8 B ╰╴test

size:     807
capacity: 1207

1.207 kB 100.00% ⏺: readme::main::Struct<readme::main::TestEnum, readme::main::Data<alloc::vec::Vec<u8>>>
   16  B   1.33% ├╴a: readme::main::TestEnum
                 │ ├╴Variant: Unnamed
    8  B   0.66% │ ├╴0: usize
    1  B   0.08% │ ╰╴1: u8 [6B]
1.183 kB  98.01% ├╴b: readme::main::Data<alloc::vec::Vec<u8>>
  724  B  59.98% │ ├╴a: alloc::vec::Vec<u8>
  424  B  35.13% │ ├╴b: alloc::vec::Vec<i32>
   35  B   2.90% │ ╰╴c: (u8, alloc::string::String)
    1  B   0.08% │   ├╴0: u8 [7B]
   27  B   2.24% │   ╰╴1: alloc::string::String
    8  B   0.66% ╰╴test: isize

size:     807
capacity: 1207

807 B ⏺
783 B ├╴b
724 B │ ├╴a
 24 B │ ├╴b
 35 B │ ╰╴c
  1 B │   ├╴0 [7B]
 27 B │   ╰╴1
 16 B ├╴a
      │ ├╴Variant: Unnamed
  1 B │ ├╴1 [6B]
  8 B │ ╰╴0
  8 B ╰╴test

Caveats

• We support out-of-the-box most basic types, and tuples up to size ten. The derive macros MemSize/MemDbg will generate implementations for structs and enums whose fields implement the associated interface: if this is not the case (e.g., because of the orphan rule) one can implement the traits manually.

• Computation of the size of arrays, slices, and vectors will be performed by iterating over their elements unless the type is a copy type that does not contain non-'static references and it is declared as such using the attribute #[copy_type]. See CopyType for more details.

• The content of vectors and slices is not expanded recursively as the output might be too complex; this might change in the future (e.g., via a flag) should interesting use cases arise.

• BTreeMap, and BTreeSet, are not currently supported as we still have to figure out a way to precisely measure their memory size and capacity.

mem_dbg-rs's Issues

Use offset_of to implement padding reporting

The stabilization of offset_of makes it possible to

• Print fields in the order in which they are actually stored
• Print padding information
• In the future, have a simpler repr_hash; however, this might invalidate serialized data unless we find a simple way to keep both systems working using version numbers.

Expose half feature

Can you expose the half feature in the cargo.toml?