[ https://issues.apache.org/jira/browse/CASSANDRA-6689?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13918785#comment-13918785 ]
Pavel Yaskevich edited comment on CASSANDRA-6689 at 3/4/14 12:17 AM: --------------------------------------------------------------------- bq. If I understand you correctly, you mean to deal with the situation where we want to flush memtables quick enough that the single cleaner thread may become the bottleneck? In which case, I don't think this should ever be a problem: the work done by the cleaner itself is very minimal unless it's got some GC to do, in which case it is potentially non-trivial, but still unlikely to be a bottleneck. There might be a case for handing off the GC candidate selection to another thread, except that we expect it to certainly be quicker than any flush, and we want to wait for it to complete before deciding if flushing is actually worthwhile, so I'm not sure it buys us much. Maybe we could do this and continue to check the memory state isn't worsening, and if it is just force a flush without waiting for the GC, but it probably is too unlikely to be of benefit to be worth the extra complexity. That is, unless you meant some other benefit I missed? I can image the situation when we frequently switch memtables which, in current code, starts a new thread, so I wonder how quickly we would be able to see a slowdown affect caused by that + potentially OOM on stack space... bq. The pool is global; a per-cf pool would have the issue of poor sharing of resources, though, no? I would probably like to see us move to allocators surviving across memtable replacement, so that we could then flatten allocator group with allocator, but I'd prefer to save that for a later date. I'm thinking maybe it would lessen internal complexity if we do (sort-of) arena allocation inside instead of having groups and sub pools? bq. Not really sure what you're suggesting here, could you elaborate? It's an idea so people can just ignore it but what I meant there is maybe it would be a better first step for us to consider making all of the operations that we do more self-contained so we can track exactly where it started and ends (by operation I mean all types of reads, writes), which could potentially make allocator job easier as we would be able to allocate memory per operation and get rid of RefAction as operation wouldn't be able to leak references to outside... was (Author: xedin): bq. If I understand you correctly, you mean to deal with the situation where we want to flush memtables quick enough that the single cleaner thread may become the bottleneck? In which case, I don't think this should ever be a problem: the work done by the cleaner itself is very minimal unless it's got some GC to do, in which case it is potentially non-trivial, but still unlikely to be a bottleneck. There might be a case for handing off the GC candidate selection to another thread, except that we expect it to certainly be quicker than any flush, and we want to wait for it to complete before deciding if flushing is actually worthwhile, so I'm not sure it buys us much. Maybe we could do this and continue to check the memory state isn't worsening, and if it is just force a flush without waiting for the GC, but it probably is too unlikely to be of benefit to be worth the extra complexity. That is, unless you meant some other benefit I missed? I can image the situation when we frequently switch memtables which, in current code, starts a new thread, so I wonder how quickly we would be able to see a slowdown affect caused by that + potentially OOM on stack space... bq. The pool is global; a per-cf pool would have the issue of poor sharing of resources, though, no? I would probably like to see us move to allocators surviving across memtable replacement, so that we could then flatten allocator group with allocator, but I'd prefer to save that for a later date. I'm thinking maybe it would lessen internal complexity if we do (sort-of) arena allocation inside instead of having groups and sub pools? bq. Not really sure what you're suggesting here, could you elaborate? It's just an idea so people can just ignore it but what I meant there is maybe it would be a better first step for us to consider making all of the operations that we do more self-contained so we can track exactly where it started and ends (by operation I mean all types of reads, writes), which could potentially make allocator job easier as we would be able to container per operation memory... > Partially Off Heap Memtables > ---------------------------- > > Key: CASSANDRA-6689 > URL: https://issues.apache.org/jira/browse/CASSANDRA-6689 > Project: Cassandra > Issue Type: New Feature > Components: Core > Reporter: Benedict > Assignee: Benedict > Fix For: 2.1 beta2 > > > Move the contents of ByteBuffers off-heap for records written to a memtable. > (See comments for details) -- This message was sent by Atlassian JIRA (v6.2#6252)