[
https://issues.apache.org/jira/browse/CASSANDRA-7546?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
graham sanderson updated CASSANDRA-7546:
----------------------------------------
Attachment: 7546.20_2.txt
Yes, I agree with you; simpler is better - I didn't like the fact that it was
hard to disentangle my code.
So I have attached a different patch 7546.20_2.txt, which is about the same as
yours except that it tracks a very minimal state heuristic to avoid the first
unlocked loop when it thinks it is contended... this helps under load where it
attempts about 1.11 loop per mutation in the heavily contented case vs 1.93
with 7546.20.txt (or about 18 without any patch at all).
It turns out to be rather hard to measure this with cassandra-stress in
isolation - basically the looping causes very high memory allocation rates
which may or may not be a problem based on the current state of memory in the
JVM (e.g. currently resident recent memtables) and other activity. Also, I may
open a separate ticket to allow batch (thrift and maybe CQL) inserts in
cassandra-stress to bump up the load a little more.
That said, clearly this race is undesirable, and we have certainly observed its
effects in practice in the hinting case (which is doubly unfortunate since you
are already now doing more work per node than you thought, and you are going
down a contention path you weren't expecting based on your own partitioning).
Note when it goes wrong it goes really wrong!
So, I think settling for a patch with minimal code or uncontended path impact,
that caps us nicely is the right way to go
> AtomicSortedColumns.addAllWithSizeDelta has a spin loop that allocates memory
> -----------------------------------------------------------------------------
>
> Key: CASSANDRA-7546
> URL: https://issues.apache.org/jira/browse/CASSANDRA-7546
> Project: Cassandra
> Issue Type: Bug
> Components: Core
> Reporter: graham sanderson
> Assignee: graham sanderson
> Attachments: 7546.20.txt, 7546.20_2.txt, 7546.20_alt.txt,
> suggestion1.txt, suggestion1_21.txt
>
>
> In order to preserve atomicity, this code attempts to read, clone/update,
> then CAS the state of the partition.
> Under heavy contention for updating a single partition this can cause some
> fairly staggering memory growth (the more cores on your machine the worst it
> gets).
> Whilst many usage patterns don't do highly concurrent updates to the same
> partition, hinting today, does, and in this case wild (order(s) of magnitude
> more than expected) memory allocation rates can be seen (especially when the
> updates being hinted are small updates to different partitions which can
> happen very fast on their own) - see CASSANDRA-7545
> It would be best to eliminate/reduce/limit the spinning memory allocation
> whilst not slowing down the very common un-contended case.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
