cassandra-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Randy Fradin (JIRA)" <>
Subject [jira] [Commented] (CASSANDRA-12965) StreamReceiveTask causing high CPU utilization during repair
Date Tue, 20 Dec 2016 21:26:58 GMT


Randy Fradin commented on CASSANDRA-12965:

Understood on not fixing in 2.1- will still be nice to see that it's fixed for when we upgrade.
Here's the info you asked for:

- This happened more than once. We had a data center's worth of nodes down for a long period
of time (longer than the hinted handoff window) before this happened so I am assuming that
caused more ranges to be out of sync than usual before this repair run. The tables were not
particularly big (a few GB total at most) so it could not have been a large volume of data
that needed be synced, but nevertheless it resulted in thousands of SSTables being created
on the nodes that had been down for a set of tables that normally have ~20ish SSTables. After
killing repair, running it again would yield the same result. We avoided running repair on
those particular tables until we could figure out what to do. The large number of SSTables
caused its own problems that we worked around, but separate from that we had this CPU problem
resulting from all the streaming sessions that created the SSTables.
- We run full (non-incremental) repair with the -pr and -par options. Each run is always for
a specific table.
- We have around 400 tables in this cluster with varying RFs, but the RF for the tables that
were causing the issue is 3 per data center across 4 data centers. There are 24 nodes total
in the cluster and each node has 256 vnodes.
- Yes we have our own repair coordinator that's currently configured to run up to 8 repairs
at the same time across the cluster.

> StreamReceiveTask causing high CPU utilization during repair
> ------------------------------------------------------------
>                 Key: CASSANDRA-12965
>                 URL:
>             Project: Cassandra
>          Issue Type: Bug
>            Reporter: Randy Fradin
> During a full repair run, I observed one node in my cluster using 100% cpu (100% of all
cores on a 48-core machine). When I took a stack trace I found exactly 48 running StreamReceiveTask
threads. Each was in the same block of code in StreamReceiveTask.OnCompletionRunnable:
> {noformat}
> "StreamReceiveTask:8077" #1511134 daemon prio=5 os_prio=0 tid=0x00007f01520a8800 nid=0x6e77
runnable [0x00007f020dfae000]
>    java.lang.Thread.State: RUNNABLE
>         at java.util.ComparableTimSort.binarySort(
>         at java.util.ComparableTimSort.sort(
>         at java.util.Arrays.sort(
>         at java.util.Arrays.sort(
>         at java.util.ArrayList.sort(
>         at java.util.Collections.sort(
>         at org.apache.cassandra.utils.IntervalTree$IntervalNode.<init>(
>         at org.apache.cassandra.utils.IntervalTree$IntervalNode.<init>(
>         at org.apache.cassandra.utils.IntervalTree.<init>(
>         at org.apache.cassandra.db.DataTracker$SSTableIntervalTree.<init>(
>         at org.apache.cassandra.db.DataTracker$SSTableIntervalTree.<init>(
>         at org.apache.cassandra.db.DataTracker.buildIntervalTree(
>         at org.apache.cassandra.db.DataTracker$View.replace(
>         at org.apache.cassandra.db.DataTracker.addSSTablesToTracker(
>         at org.apache.cassandra.db.DataTracker.addSSTables(
>         at org.apache.cassandra.db.ColumnFamilyStore.addSSTables(
>         at org.apache.cassandra.streaming.StreamReceiveTask$
>         at java.util.concurrent.Executors$
>         at
>         at java.util.concurrent.ThreadPoolExecutor.runWorker(
>         at java.util.concurrent.ThreadPoolExecutor$
>         at
> {noformat}
> All 48 threads were in ColumnFamilyStore.addSSTables(), and specifically in the IntervalNode
constructor called from the IntervalTree constructor.
> It stayed this way for maybe an hour before we restarted the node. The repair was also
generating thousands (20,000+) of tiny SSTables in a table that previously had just 20.
> I don't know enough about SSTables and ColumnFamilyStore to know if all this CPU work
is necessary or a bug, but I did notice that these tasks are run on a thread pool constructed
in, so perhaps this pool should have a thread count max less than the
number of processors on the machine, at least for machines with a lot of processors. Any reason
not to do that? Any ideas for a reasonable # or formula to cap the thread count?
> Some additional info: We have never run incremental repair on this cluster, so that is
not a factor. All our tables use LCS. Unfortunately I don't have the log files from the period

This message was sent by Atlassian JIRA

View raw message