Currently BlockCacheIO caches byte arrays in equal sizes of blockSize/minIOSeekSize each. This provides faster HashMap inserts and lookups since keys themselves are just Long values (unique and no hash conflicts) but
It has the cost of needing to join multiple Array[Byte] if the required data is greater than minIOSeekSize.
Need multiple ConcurrentHashMap lookups if the required data is greater than minIOSeekSize.
Synchronised - java.nio.ScatteringByteChannel already provides an API which slices the data into smaller Array[Byte]s of minIOSeekSize each (so not in-memory slicing/copying is required) but this API still needs synchronised (blocking) access since ScatteringByteChannel mutates the position in the FileChannel.
Synchronised can be removed by not using ScatteringByteChannel and slicing data manually in-memory but this creates a lot more on-heap allocations which costs GC time.
Task
Another BlockCacheSkipListIO based on ConcurrentSkipList is required which allows caching randomly sized data. This will not need synchronised access and will provide easier caching but in comparison to ConcurrentHashMap, inserts and lookups will be slower which in some cases might not be noticeable since multiple lookups to join multiple byte arrays (in ConcurrentHashMap's cases) might not be required.
Configuration
Either one of the caching strategy should be configurable.
Overview
Currently
BlockCacheIOcaches byte arrays in equal sizes ofblockSize/minIOSeekSizeeach. This provides fasterHashMapinserts and lookups since keys themselves are justLongvalues (unique and no hash conflicts) butArray[Byte]if the required data is greater thanminIOSeekSize.ConcurrentHashMaplookups if the required data is greater thanminIOSeekSize.Synchronised-java.nio.ScatteringByteChannelalready provides an API which slices the data into smallerArray[Byte]s ofminIOSeekSizeeach (so not in-memory slicing/copying is required) but this API still needssynchronised(blocking) access sinceScatteringByteChannelmutates thepositionin theFileChannel.Synchronisedcan be removed by not usingScatteringByteChanneland slicing data manually in-memory but this creates a lot more on-heap allocations which costs GC time.Task
Another
BlockCacheSkipListIObased onConcurrentSkipListis required which allows caching randomly sized data. This will not needsynchronisedaccess and will provide easier caching but in comparison toConcurrentHashMap, inserts and lookups will be slower which in some cases might not be noticeable since multiple lookups to join multiple byte arrays (inConcurrentHashMap's cases) might not be required.Configuration
Either one of the caching strategy should be configurable.