[ https://issues.apache.org/jira/browse/CASSANDRA-11870?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15344121#comment-15344121 ]
Robert Stupp commented on CASSANDRA-11870: ------------------------------------------ It's a bit difficult to tell whether it's experienced in the wild, since before 3.x Netty buffers were on-heap and C* only had somewhat infrequent direct memory allocations. However, with 3.0 and especially 3.6, we have more frequent off-heap memory allocations. At least from my point of view, the majority is still on 2.1. What I intend to tackle is to prevent the (ugly) call to {{java.nio.Bits#reserveMemory}} and also the global linked queue via {{sun.misc.Cleaner}}, which has static synchronized methods {{add}} + {{remove}}. My objection about {{ByteBuffer.allocateDirect()}} in general is that if one thread runs into an _out-of-direct-memory_ situations, other thread will likely run into that situation, too - and that will cause one CMS-GC ({{System.gc()}}) per thread. However, I agree that maintaining a separate {{max_direct_memory_in_mb}} config variable is probably not the best way. A better option could be to use the already existing options that define off-heap memory bounds like {{file_cache_size_in_mb}}, {{memtable_offheap_space_in_mb}}, {{index_summary_capacity_in_mb}}, sum them and use that value. WDYT? I'm not sure, whether the following is a good idea: while testing OHC a while ago with >30G of small allocations, it worked really nice (OHC does not use BB.allocateDirect - that time I just realized that it's slow). What I'm thinking of, is to not pool off-heap memory at all (and just keep track of the total allocated amount, as currently in the patch) and leave alloc()/free() up to the OS or jemalloc. Your idea to centralize BB allocations in BufferPool sounds good! > Consider allocating direct buffers bypassing ByteBuffer.allocateDirect > ---------------------------------------------------------------------- > > Key: CASSANDRA-11870 > URL: https://issues.apache.org/jira/browse/CASSANDRA-11870 > Project: Cassandra > Issue Type: Improvement > Reporter: Robert Stupp > Assignee: Robert Stupp > Priority: Minor > Fix For: 3.x > > > As outlined in CASSANDRA-11818, {{ByteBuffer.allocateDirect}} uses > {{Bits.reserveMemory}}, which is there to respect the JVM setting > {{-XX:MaxDirectMemorySize=...}}. > {{Bits.reserveMemory}} first tries an "optimistic" {{tryReserveMemory}} and > exits immediately on success. However, if that somehow doesn't succeed, it > triggers a {{System.gc()}}, which is bad IMO (however, kind of how direct > buffers work in Java). After that GC it sleeps and tries to reserve the > memory up to 9 times - up to 511 ms - and then throws > {{OutOfMemoryError("Direct buffer memory")}}. > This is unnecessary for us since we always immediately "free" direct buffers > as soon as we no longer need them. > Proposal: Manage direct-memory reservations in our own code and skip > {{Bits.reserveMemory}} that way. > (However, Netty direct buffers are not under our control.) -- This message was sent by Atlassian JIRA (v6.3.4#6332)