OAK-11932: Segment prefetching for CachingSegmentReader #2513
Conversation
reschke
changed the title
| return out; | ||
| } | ||
|
|
||
| private void schedulePrefetch(long msb, long lsb, Buffer buffer) { |
There was a problem hiding this comment.
@nfsantos made a good remark that this method, from the start, should be executed in a separate thread context, so it does not impact execution time for the thread invoking CachingSegmentArchiveReader#readSegment
There was a problem hiding this comment.
Indeed, my tests with #2519, which implements a similar mechanism, but on a different level of abstraction, showed that dispatching from the caller thread hurts performance.
I even opted to trigger preloading only from within the load-callback, so preloading is completely off the critical path for cache hits.
Of course that's a trade-off, as it requires the caller thread to load one segment before any preloading happens.
| return out; | ||
| } | ||
|
|
||
| private void schedulePrefetch(long msb, long lsb, Buffer buffer) { |
There was a problem hiding this comment.
Indeed, my tests with #2519, which implements a similar mechanism, but on a different level of abstraction, showed that dispatching from the caller thread hurts performance.
I even opted to trigger preloading only from within the load-callback, so preloading is completely off the critical path for cache hits.
Of course that's a trade-off, as it requires the caller thread to load one segment before any preloading happens.
| } | ||
|
|
||
| // Drop prefetch if already in progress for this segment | ||
| boolean registered = inFlightPrefetch.add(ref); |
There was a problem hiding this comment.
Nice idea! I missed this trick so far in #2519 🙂
| List<UUID> refs = extractReferences(buffer); | ||
| int limit = Math.min(refs.size(), prefetchMaxRefs); | ||
| for (int i = 0; i < limit; i++) { |
There was a problem hiding this comment.
You are getting a list with all the references but then potentially iterate over only a subset of them. You could save some work by extracting only the references that will be prefetched, this could also avoid allocating an array with all the refs. Using streams with limit()may be an easy way of implementing this optimization, as streams are computed lazily.
| if (persistentCache.containsSegment(rMsb, rLsb)) { | ||
| continue; | ||
| } |
There was a problem hiding this comment.
Maybe it's better to do this check in the worker thread, just before the segment is downloaded? The task that is scheduled to download the segment might not execute for a while, if the worker pool is busy, so from this point until the actual download, the segment might have been added to the cache. Or we can leave the check here and do another one before trying to download.
It would also be better to have a mechanism similar to the Guava loading cache, where thread requesting a segment that is not in the cache but is being downloaded will block waiting for the first download to complete, instead of starting a new download. This would avoid duplicate downloads.
|
As indicated in #2569, this approach is limited in that a |
3 participants
https://issues.apache.org/jira/browse/OAK-11932