#cache #memory-mapped #mmap #low-level #offset #byte-offset #key

mmap-cache

A low-level, read-only, memory-mapped cache

1 unstable release

0.1.0 May 7, 2022

#290 in Caching

MIT/Apache

25KB
392 lines

mmap-cache

A low-level API for a memory-mapped cache of a read-only key-value store.

Design

The Cache index is an fst::Map, which maps from arbitrary byte sequences to [u64]. We use the [u64] as byte offsets into a memory-mapped file that stores arbitrary values. This read-only cache is very compact and simple to construct, requiring only constant memory for serializing arbitrarily large maps.

By using Finite State Transducers from the [fst] crate, we get a highly compressed mapping that performs key --> offset lookups in O(key length) time.

Example

use mmap_cache::{FileBuilder, MmapCache};

const INDEX_PATH: &str = "/tmp/mmap_cache_index";
const VALUES_PATH: &str = "/tmp/mmap_cache_values";

// Serialize to files. As required by the finite state transducer (FST) builder,
// keys must be provided in sorted (lexicographical) order.
let mut builder = FileBuilder::create_files(INDEX_PATH, VALUES_PATH)?;
builder.insert(b"abc", b"def")?;
builder.insert(b"foo", b"bar")?;
builder.finish()?;

// Map the files into memory.
let cache = unsafe { MmapCache::map_paths(INDEX_PATH, VALUES_PATH) }?;
let value = unsafe { cache.get_transmuted_value(b"foo") };
assert_eq!(value, Some(b"bar"));

IO Concurrency

When using memmap2 on a large file, it's likely that accessing values from the cache will cause the thread to block in the operating system scheduler while the page cache is filled from the file system. To achieve concurrency IO up to some maximum concurrency N, you could dispatch your IOs in a thread pool of N threads.

License: MIT OR Apache 2.0

Dependencies

~1.5MB