Peter Xie created CASSANDRA-14653:
-------------------------------------
Summary: The performance of "NonPeriodicTasks" pools defined in
class ScheduledExecutors is low
Key: CASSANDRA-14653
URL: https://issues.apache.org/jira/browse/CASSANDRA-14653
Project: Cassandra
Issue Type: Improvement
Components: Compaction
Environment: Cassandra nodes :
3 nodes, 330G physical memory per node , and four data directory (ssd) per
node.
Reporter: Peter Xie
We use cassandra as backend storage for Janusgraph. when we loading huge data
(~2 billion vertex, ~10 billion edges), we met an problems.
At first, we use STCS as compaction strategy , but met below exception. we
checked the value of "max memory lock" is unlimited and "map count" is 1
million, these values should enough for loading. last , we found this problem
is caused by the virtual memory are all cosumed by cassandra. So not
additional virtual memory can be used by compaction task , and below exception
is thrower out.
{quote}ERROR [CompactionExecutor:267] 2018-08-09 02:28:40,952
JVMStabilityInspector.javv
a:74 - OutOfMemory error letting the JVM handle the error:
java.lang.OutOfMemoryError: Map failed
{quote}
So, we change compaction strategy to LCS, this change seems can resolve the
virtual memory problem. But we found another problem : Many sstables which has
been compacted are still retained on disk, at last these old sstable consume so
many disk space, it's causing no enough disk for saving real data. we found
that so many files like "mc_txn_compaction_xxx.log" are created under the data
directory.
After some times' investigaton, we found that this problem is caused by
"NonPeriodicTasks" thread pools. this pools is always using only one thread
for processing clean task and compaction. this thread pool is instanced with
class DebuggableScheduledThreadPoolExecutor,
and DebuggableScheduledThreadPoolExecutor is inherit from class
ScheduledThreadPoolExecutor.
By reading the code of class DebuggableScheduledThreadPoolExecutor, found
DebuggableScheduledThreadPoolExecutor is using an unbound task queue, and core
pool size is 1. Why here use the unbound queue for queuing submitted tasks?
If we using unbound queue, the thread pool wouldn't increasing thread even
there so many task are blocked in queue, because unbound queue never would be
full. I think here should use bound queue, so when task is heavily, more
threads would created for processing them.
{quote}public DebuggableScheduledThreadPoolExecutor(int corePoolSize, String
threadPoolName, int priority)
{
super(corePoolSize, new NamedThreadFactory(threadPoolName, priority));
setRejectedExecutionHandler(rejectedExecutionHandler);
}
public ScheduledThreadPoolExecutor(int corePoolSize,
ThreadFactory threadFactory) {
super(corePoolSize, Integer.MAX_VALUE, 0, NANOSECONDS,
new DelayedWorkQueue(), threadFactory);
}
{quote}
Below is the case about clean task after compaction. there nearly 3 hours
delay for removing file "mc-56525".
{quote}
TRACE [CompactionExecutor:81] 2018-08-16 21:22:29,664
LifecycleTransaction.java:363 - Staging for obsolescence
BigTableReader(path='/sdb/data/test_2/edgestore-365b0b70a05911e8806001ebe60a5ce7/mc-56525-big-Data.db')
..........
TRACE [CompactionExecutor:81] 2018-08-16 21:22:41,162 Tracker.java:165 -
removing
/sdb/data/test_2/edgestore-365b0b70a05911e8806001ebe60a5ce7/mc-56525-big from
list of files tracked for test_2.edgestore
............
TRACE [NonPeriodicTasks:1] 2018-08-17 00:28:47,179 SSTableReader.java:2175 -
Async instance tidier for
/sdb/data/test_2/edgestore-365b0b70a05911e8806001ebe60a5ce7/mc-56525-big,
before barrier
TRACE [NonPeriodicTasks:1] 2018-08-17 00:28:47,180 SSTableReader.java:2181 -
Async instance tidier for
/sdb/data/test_2/edgestore-365b0b70a05911e8806001ebe60a5ce7/mc-56525-big, after
barrier
TRACE [NonPeriodicTasks:1] 2018-08-17 00:28:47,182 SSTableReader.java:2196 -
Async instance tidier for
/sdb/data/test_2/edgestore-365b0b70a05911e8806001ebe60a5ce7/mc-56525-big,
completed
{quote}
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
