Erlang 2Q NIF cache.
This is an implementation of the 2Q Cache Management algorithm (http://www.inf.fu-berlin.de/lehre/WS10/DBS-Tech/Reader/2QBufferManagement.pdf) as an Erlang NIF.
2Q is a refinement of the classic LRU (Least-Recently-Used) cache management algorithm that achieves better hit rates with many common workloads -- especially those that benefit from rejection of sequential scans. In the worst case, it performs no worse than plain LRU and still retains much of its simplicity.
This implementation's primary goals are:
- a very simple to use API, easy to integrate into your project
- high performance when having high hit rate (ie, hits being fast is preferred over misses being fast)
Cache hits can be zero-copy (using resource binaries) and the updates to the cache structure are deferred to a background thread to avoid blocking the Erlang VM. Benchmarks welcome!
-
Add e2qc as a
rebar.config
dependency in your project:{deps, [ {e2qc, ".*", {git, "git://github.com/arekinath/e2qc.git", "HEAD"}} ]}
-
Use it! Wrap your slow processing that you want to cache in a call to
e2qc:cache
:some_function(Input) -> do_slow_thing(Input). becomes some_function(Input) -> e2qc:cache(slow_thing, Input, fun() -> do_slow_thing(Input) end).
It's really that simple. Each "cache" is named by a unique atom (in this case we've used a cache called slow_thing
). You don't need to explicitly create or configure the cache before using it -- it will be created on the first use. The default configuration will cache up to 4MB of data.
You can use the same cache from any number of Erlang processes at once on the same node (and it will be just one shared cache). No passing around handles or pids for the cache, and no extra setup required.
If you want to adjust the size
of the cache, or set a different Q1 ratio
(see the paper on the 2Q algorithm for details; the default is 0.3 or 30%), use the e2qc:setup
function:
ok = e2qc:setup(slow_thing, [{size, 16*1024*1024}, {ratio, 0.4}]).
Put this in your startup procedure somewhere and it will configure the slow_thing
cache with a 16MB size instead of the default 4MB, and a Q1 ratio of 0.4 (the value for the target size of Q1 during eviction will be 6.4MB).
Currently, if you make a call to e2qc:setup/2
with a size that is smaller than the default, the call has to happen before the cache is used for the first time (otherwise it will throw an error). Fixing this is an open TODO.
There is also support now for timed eviction or expiry of entries. This is also easy to use -- just call e2qc:cache/4
instead of e2qc:cache/3
:
some_function(Input) ->
e2qc:cache(slow_thing, Input, 30, fun() ->
do_slow_thing(Input)
end).
This will keep the cache entries for a maximum of 30 seconds and then automatically evict them. The argument is in seconds and not milliseconds (like most other Erlang timers), because to keep the overhead of expiry to a minimum, e2qc sacrifices precision on this parameter. If you need high-precision expiry times, you should use a normal timer and a call to e2qc:evict/2
.
The e2qc:stats/1
function is useful if you want to know how your cache is doing:
=> e2qc:stats(slow_thing).
[{hits,5674},{misses,11},{q1size,280},{q2size,735}]
If you know that an entry in the cache is stale or needs to be evicted, you can use the e2qc:evict/2
function to clear it out:
e2qc:evict(slow_thing, OldInput)
Now the next attempt to look for OldInput
will miss and be re-calculated.
You can also destroy an entire cache if you wish, using e2qc:teardown/1
. This will destroy the cache and all of its entries entirely (but note that if another call attempts to use it afterwards, it will be re-created implicitly with default settings).
- Shrinking with
e2qc:setup/2
after cache has already started