Page-based cache recycling. #4559
Conversation
| } | ||
| context.queryResult().aggregations(new InternalAggregations(aggregations)); | ||
| } finally { | ||
| for (Aggregator aggregator : aggregators) { |
There was a problem hiding this comment.
do we have some utils for Releasable like IOUtils.close(Closeable) that ensures that all of them are released?
ghost
assigned 
| public static ByteSizeValue parseBytesSizeValueOrHeapRatio(String sValue) { | ||
| if (sValue.endsWith("%")) { | ||
| double percent = Double.parseDouble(sValue.substring(0, sValue.length() - 1)); | ||
| return new ByteSizeValue((long) ((percent / 100) * JvmInfo.jvmInfo().getMem().getHeapMax().bytes()), ByteSizeUnit.BYTES); |
There was a problem hiding this comment.
I think putting the heap knowledge in here is the wrong place encapsulation wise? Maybe add to Settings a getAsMemory, that will do this?
|
New commit pushed:
|
|
|
||
| private static int maximumSearchThreadPoolSize(ThreadPool threadPool) { | ||
| final Executor executor = threadPool.executor(ThreadPool.Names.SEARCH); | ||
| return ((ThreadPoolExecutor) executor).getMaximumPoolSize(); |
There was a problem hiding this comment.
can we use the ThreadPool#info API here, so we don't have to cast to ThreadPoolExecutor, and be able to get the max back through the info class.
| @Inject | ||
| public MockPageCacheRecycler(Settings settings, ThreadPool threadPool) { | ||
| super(settings, threadPool); | ||
| random = new Random(0); |
There was a problem hiding this comment.
the settings should get a random seed per index maybe I should pass on the node seed from the test cluster as well so you can get it?
|
added some smallish comments - looks really good though. Lets get this in soon! |
Refactor cache recycling so that it only caches large arrays (pages) that can
later be used to build more complex data-structures such as hash tables.
- QueueRecycler now takes a limit like other non-trivial recyclers.
- New PageCacheRecycler (inspired of CacheRecycler) has the ability to cache
byte[], int[], long[], double[] or Object[] arrays using a fixed amount of
memory (either globally or per-thread depending on the Recycler impl, eg.
queue is global while thread_local is per-thread).
- Paged arrays in o.e.common.util can now optionally take a PageCacheRecycler
to reuse existing pages.
- All aggregators' data-structures now use PageCacheRecycler:
- for all arrays (counts, mins, maxes, ...)
- LongHash can now take a PageCacheRecycler
- there is a new BytesRefHash (inspired from Lucene but quite different,
still; for instance it cheats on BytesRef comparisons by using Unsafe)
that also takes a PageCacheRecycler
Close elastic#4557
|
Thanks for the comments, @s1monw. I rebased and did the changes you suggested:
|
|
LGTM +1 to push |
Here is an attempt to refactor cache recycling so that it only caches large
arrays (pages) that can later be used to build more complex data-structures
such as hash tables.
byte[], int[], long[], double[] or Object[] arrays using a fixed amount of
memory (either globally or per-thread depending on the Recycler impl, eg.
queue is global while thread_local is per-thread).
to reuse existing pages.
still; for instance it cheats on BytesRef comparisons by using Unsafe)
that also takes a PageCacheRecycler
Close #4557