Thread: Postgresql server gets stuck at low load

De: "Krzysztof Olszewski" <kolszew73@gmail.com>
Para: pgsql-performance@lists.postgresql.org
Enviadas: Sexta-feira, 5 de junho de 2020 7:07:02
Assunto: Postgresql server gets stuck at low load

I have problem with one of my Postgres production server. Server works fine almost always, but sometimes without any increase of transactions or statements amount, machine gets stuck. Cores goes up to 100%, load up to 160%. When it happens then there are problems with connect to database and even it will succeed, simple queries works several seconds instead of milliseconds.Problem sometimes stops after a period a time (e.g. 35 min), sometimes we must restart Postgres, Linux, or even KVM (which exists as virtualization host).

My hardware56 cores (Intel Core Processor (Skylake, IBRS))400 GB RAMRAID10 with about 40k IOPS

Os

CentOS Linux release 7.7.1908

kernel 3.10.0-1062.18.1.el7.x86_64 Databasesize 100 GB (entirely fit in memory :) )server_version 10.12effective_cache_size 192000 MBmaintenance_work_mem 2048 MBmax_connections 150 shared_buffers 64000 MBwork_mem 96 MBOn normal state, i have about 500 tps, 5% usage of cores, about 3% of load, whole database fits in memory, no reads from disk, only writes on about 500 IOPS level, sometimes in spikes on 1500 IOPS level, but on this hardware there is no problem with this values (no iowaits on cores). In normal state this machine does "nothing". Connections to database are created by two app servers based on Java, through connection pools, so connections count is limited by configuration of pools and max is 120, is lower value than in Postgres configuration (150). On normal state there is about 20 connections, when stuck goes into max (120).In correlation with stucks i see informations in kernel log aboutNMI watchdog: BUG: soft lockup - CPU#25 stuck for 23s! [postmaster:33935]but i don't know this is reason or effect of problemI made investigation with pgBadger and ... nothing strange happens, just normal statements Any ideas? Thanks,

Kris

Hi Krzysztof!

I would enable pg_stat_statements extension and check if there are long running queries that should be quick.

pá 5. 6. 2020 v 12:07 odesílatel Krzysztof Olszewski <kolszew73@gmail.com> napsal:

I have problem with one of my Postgres production server. Server works fine almost always, but sometimes without any increase of transactions or statements amount, machine gets stuck. Cores goes up to 100%, load up to 160%. When it happens then there are problems with connect to database and even it will succeed, simple queries works several seconds instead of milliseconds.Problem sometimes stops after a period a time (e.g. 35 min), sometimes we must restart Postgres, Linux, or even KVM (which exists as virtualization host).

My hardware
56 cores (Intel Core Processor (Skylake, IBRS))
400 GB RAM
RAID10 with about 40k IOPS

Os

CentOS Linux release 7.7.1908

kernel 3.10.0-1062.18.1.el7.x86_64

Databasesize 100 GB (entirely fit in memory :) )
server_version 10.12
effective_cache_size 192000 MB
maintenance_work_mem 2048 MB
max_connections 150
shared_buffers 64000 MB
work_mem 96 MB

On normal state, i have about 500 tps, 5% usage of cores, about 3% of load, whole database fits in memory, no reads from disk, only writes on about 500 IOPS level, sometimes in spikes on 1500 IOPS level, but on this hardware there is no problem with this values (no iowaits on cores). In normal state this machine does "nothing". Connections to database are created by two app servers based on Java, through connection pools, so connections count is limited by configuration of pools and max is 120, is lower value than in Postgres configuration (150). On normal state there is about 20 connections, when stuck goes into max (120).

In correlation with stucks i see informations in kernel log about
NMI watchdog: BUG: soft lockup - CPU#25 stuck for 23s! [postmaster:33935]
but i don't know this is reason or effect of problem
I made investigation with pgBadger and ... nothing strange happens, just normal statements

Any ideas?

you can try to install perf + debug symbols for postgres. When you will have this problem again run "perf top". You can see what routines eat your CPU.

Maybe it can be a spinlock problem

https://www.postgresql.org/message-id/CAHyXU0yAsVxoab2PcyoCuPjqymtnaE93v7bN4ctv2aNi92fefA%40mail.gmail.com

Can be interesting a reply on Merlin's question from mail/.

cat /sys/kernel/mm/redhat_transparent_hugepage/enabled
cat /sys/kernel/mm/redhat_transparent_hugepage/defrag

Regards

Pavel

 

Thanks,

Kris

Hi,

On Fri, Jun 5, 2020 at 7:07 AM Krzysztof Olszewski <kolszew73@gmail.com> wrote:

I have problem with one of my Postgres production server. Server works fine almost always, but sometimes without any increase of transactions or statements amount, machine gets stuck. Cores goes up to 100%, load up to 160%. When it happens then there are problems with connect to database and even it will succeed, simple queries works several seconds instead of milliseconds.Problem sometimes stops after a period a time (e.g. 35 min), sometimes we must restart Postgres, Linux, or even KVM (which exists as virtualization host).

My hardware
56 cores (Intel Core Processor (Skylake, IBRS))
400 GB RAM
RAID10 with about 40k IOPS

Os

CentOS Linux release 7.7.1908

kernel 3.10.0-1062.18.1.el7.x86_64

Databasesize 100 GB (entirely fit in memory :) )
server_version 10.12
effective_cache_size 192000 MB
maintenance_work_mem 2048 MB
max_connections 150
shared_buffers 64000 MB
work_mem 96 MB

What is the value set to random_page_cost ? 

Set to 1 (same as default seq_page_cost) for a moment and try it. 

On normal state, i have about 500 tps, 5% usage of cores, about 3% of load, whole database fits in memory, no reads from disk, only writes on about 500 IOPS level, sometimes in spikes on 1500 IOPS level, but on this hardware there is no problem with this values (no iowaits on cores). In normal state this machine does "nothing". Connections to database are created by two app servers based on Java, through connection pools, so connections count is limited by configuration of pools and max is 120, is lower value than in Postgres configuration (150). On normal state there is about 20 connections, when stuck goes into max (120).

In correlation with stucks i see informations in kernel log about
NMI watchdog: BUG: soft lockup - CPU#25 stuck for 23s! [postmaster:33935]
but i don't know this is reason or effect of problem
I made investigation with pgBadger and ... nothing strange happens, just normal statements

Any ideas?

Thanks,

Kris

--

Regards,
Avinash Vallarapu