MultiThreadedCaches

MultiThreadedCache{K,V}() is a fast-ish, thread-safe cache.

This cache stores k=>v pairs that cache a deterministic computation. The only API into the cache is get!(): you can look up a key, and if it is not available, you can produce a value which will be added to the cache.

Accesses to the cache will look first in the per-thread cache, and then fall back to the shared thread-safe cache. Concurrent misses to the same key in the shared cache will coordinate, so that only one Task will perform the compuatation for that value, and the other Task(s) will block.

The per-thread caches have very low contention (usually only locked by that single Task), so a MultiThreadedCache{K,V} scales much better than the naive baseline Dict+ReentrantLock that you might use instead.

Some other approaches to concurrent caches include:

• Concurrent Hash Table
  • The theory in this package is to take advantage of the append-only aspect of a Cache to get some contention benefits over a conventional multi-threaded dictionary, but we're less sure in practice if this is actually true. (Once we had to make the adjustments to account for task migration, the benefits of the current design became less clear...)
• Multithread caches designed for low contention by sharding the key space, and keeping a separate lock per sub-cache. For example, having an array of e.g. 64 separate caches, sharded by a prefix of a key's hash, each with their own lock.
  • This package differs from this design by sharding the hash by Thread ID, rather than by the key space.
  • This tradeoff accepts greater data duplication in exchange for hopefully less contention.

julia> cache = MultiThreadedCache{Int, Int}(Dict(1=>2, 2=>3))
MultiThreadedCache{Int64, Int64}(Dict(2 => 3, 1 => 2))

julia> init_cache!(cache)
MultiThreadedCache{Int64, Int64}(Dict(2 => 3, 1 => 2))

julia> get!(cache, 2) do
           2+1
       end
3

julia> get!(cache, 5) do  # These accesses are safe from arbitrary threads.
           5+1
       end
6

julia> get!(cache, 5) do
           5+10
       end
6

Current benchmark results measuring scaling recorded against a baseline of a Dict() + ReentrantLock():

┌ Info: benchmark results
│   Threads.nthreads() = 1
│   time_serial = 0.03260747
│   time_parallel = 0.152217246
└   time_baseline = 0.131397639

┌ Info: benchmark results
│   Threads.nthreads() = 20
│   time_serial = 0.030174521
│   time_parallel = 0.307062643
└   time_baseline = 1.239406026

┌ Info: benchmark results
│   Threads.nthreads() = 100
│   time_serial = 0.030373533
│   time_parallel = 1.771401144
└   time_baseline = 5.397114559