[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Fwd: Re: High CPU usage on some of the nodes due to message coalesce

I don't think that root cause is related to Cassandra config, because the nodes are homogeneous and config for all of them are the same (16GB heap with default gc), also mutation counter and Native Transport counter is the same in all of the nodes, but only these 3 nodes experiencing 100% CPU usage (others have less than 20% CPU usage) 
I even decommissioned these 3 nodes from cluster and re-add them, but still the same
The cluster is OK without these 3 nodes (in a state that these nodes are decommissioned)

Sent using Zoho Mail

============ Forwarded message ============
From : Chris Lohfink <clohfink@xxxxxxxxx>
To : <user@xxxxxxxxxxxxxxxxxxxx>
Date : Sat, 20 Oct 2018 23:24:03 +0330
Subject : Re: High CPU usage on some of the nodes due to message coalesce
============ Forwarded message ============

1s young gcs are horrible and likely cause of some of your bad metrics. How large are your mutations/query results and what gc/heap settings are you using?

You can use https://github.com/aragozin/jvm-tools to see the threads generating allocation pressure and using the cpu (ttop) and what garbage is being created (hh --dead-young).

Just a shot in the dark, I would guess you have rather large mutations putting pressure on commitlog and heap. G1 with a larger heap might help in that scenario to reduce fragmentation and adjust its eden and survivor regions to the allocation rate better (but give it a bigger reserve space) but theres limits to what can help if you cant change your workload. Without more info on schema etc its hard to tell but maybe that can help give you some ideas on places to look. It could just as likely be repair coordination, wide partition reads, or compactions so need to look more at what within the app is causing the pressure to know if its possible to improve with settings or if the load your application is producing exceeds what your cluster can handle (needs more nodes).


On Oct 20, 2018, at 5:18 AM, onmstester onmstester <onmstester@xxxxxxxx.INVALID> wrote:

3 nodes in my cluster have 100% cpu usage and most of it is used by org.apache.cassandra.util.coalesceInternal and SepWorker.run?
The most active threads are the messaging-service-incomming.
Other nodes are normal, having 30 nodes, using Rack Aware strategy. with 10 rack each having 3 nodes. The problematic nodes are configured for one rack, on normal write load, system.log reports too many hint message dropped (cross node). also there are alot of parNewGc with about 700-1000ms and commit log isolated disk, is utilized about 80-90%. on startup of these 3 nodes, there are alot of "updateing topology" logs (1000s of them pending).
Using iperf, i'm sure that network is OK
checking NTPs and mutations on each node, load is balanced among the nodes.
using apache cassandra 3.11.2
I can not not figure out the root cause of the problem, although there are some obvious symptoms.

Best Regards

Sent using Zoho Mail