The WindowCache serves as a Java based "buffer cache", loading segments of a PackFile into the JVM heap prior to use. As JGit often wants to do reads of only tiny slices of a file, the WindowCache tries to smooth out these tiny reads into larger block-sized IO operations.
Whenever a cache miss occurs, {@link #load(PackFile,long)} is invoked byexactly one thread for the given (PackFile,position) key tuple. This is ensured by an array of locks, with the tuple hashed to a lock instance.
During a miss, older entries are evicted from the cache so long as {@link #isFull()} returns true.
Its too expensive during object access to be 100% accurate with a least recently used (LRU) algorithm. Strictly ordering every read is a lot of overhead that typically doesn't yield a corresponding benefit to the application.
This cache implements a loose LRU policy by randomly picking a window comprised of roughly 10% of the cache, and evicting the oldest accessed entry within that window.
Entities created by the cache are held under SoftReferences, permitting the Java runtime's garbage collector to evict entries when heap memory gets low. Most JREs implement a loose least recently used algorithm for this eviction.
The internal hash table does not expand at runtime, instead it is fixed in size at cache creation time. The internal lock table used to gate load invocations is also fixed in size.
The key tuple is passed through to methods as a pair of parameters rather than as a single Object, thus reducing the transient memory allocations of callers. It is more efficient to avoid the allocation, as we can't be 100% sure that a JIT would be able to stack-allocate a key tuple.
This cache has an implementation rule such that:
(PackFile,position) tuple.load() invocation there is exactly one {@link #createRef(PackFile,long,ByteWindow)} invocation to wrap aSoftReference around the cached entity.createRef() there will be exactly one call to {@link #clear(Ref)} to cleanup any resources associatedwith the (now expired) cached entity.Therefore, it is safe to perform resource accounting increments during the {@link #load(PackFile,long)} or{@link #createRef(PackFile,long,ByteWindow)} methods, and matchingdecrements during {@link #clear(Ref)}. Implementors may need to override {@link #createRef(PackFile,long,ByteWindow)} in order to embed additionalaccounting information into an implementation specific {@link Ref} subclass,as the cached entity may have already been evicted by the JRE's garbage collector.
To maintain higher concurrency workloads, during eviction only one thread performs the eviction work, while other threads can continue to insert new objects in parallel. This means that the cache can be temporarily over limit, especially if the nominated eviction thread is being starved relative to the other threads.
| |
| |
| |
| |
| |
| |