Poor performance after update from SLES11 SP1 to SP2

[Mark Smith]

Hardware: IBM X3650 M3 (2 x Xeon X5680 6C 3.33GHz), 96GB RAM. IBM X3524
with RAID 10 ext4 (noatime,nodiratime,data=writeback,barrier=0) volumes for
pg_xlog / data / indexes.

Software: SLES 11 SP2 3.0.58-0.6.2-default x86_64, PostgreSQL 9.0.4.
max_connections = 1500
shared_buffers = 16GB
work_mem = 64MB
maintenance_work_mem = 256MB
wal_level = archive
synchronous_commit = off
wal_buffers = 16MB
checkpoint_segments = 32
checkpoint_completion_target = 0.9
effective_cache_size = 32GB
Workload: OLTP, typically with 500+ concurrent database connections, same
Linux instance is also used as web server and application server. Far from
ideal but has worked well for 15 months.

Problem: We have been running PostgreSQL 9.0.4 on SLES11 SP1, last kernel
in use was 2.6.32-43-0.4, performance has always been great. Since updating
from SLES11 SP1 to SP2 we now experience many database 'stalls' (e.g.
normally 'instant' queries taking many seconds, any query will be slow,
just connecting to the database will be slow). We have trialled PostgreSQL
9.2.3 under SLES11 SP2 with the exact same results. During these periods
the machine is completely responsive but anything accessing the database is
extremely slow.

I have tried increasing sched_migration_cost from 500000 to 5000000 and
also tried setting sched_compat_yield to 1, neither of these appeared to
make a difference. I don't have the parameter 'sched_autogroup_enabled'.
Nothing jumps out from top/iostat/sar/pg_stat_activity however I am very
far from expert in interpreting their output

We have work underway to reduce our number of connections as although it
has always worked ok, perhaps it makes us particularly vulnerable to
kernel/scheduler changes.

I would be very grateful for any suggestions as to the best way to diagnose
the source of this problem and/or general recommendations?

[Vasilis Ventirozos]

Hello Mark,
if i was you i would start with making basic benchmarks regarding IO
performance (bonnie++ maybe?) and i would check my sysctl.conf. your
parameters look ok to me.
when these delays occur have you noticed whats causing it ? An output of
vmstat when the delays are happening would also help.


Vasilis Ventirozos

Post by Mark Smith
Hardware: IBM X3650 M3 (2 x Xeon X5680 6C 3.33GHz), 96GB RAM. IBM X3524
with RAID 10 ext4 (noatime,nodiratime,data=writeback,barrier=0) volumes for
pg_xlog / data / indexes.
Software: SLES 11 SP2 3.0.58-0.6.2-default x86_64, PostgreSQL 9.0.4.
max_connections = 1500
shared_buffers = 16GB
work_mem = 64MB
maintenance_work_mem = 256MB
wal_level = archive
synchronous_commit = off
wal_buffers = 16MB
checkpoint_segments = 32
checkpoint_completion_target = 0.9
effective_cache_size = 32GB
Workload: OLTP, typically with 500+ concurrent database connections, same
Linux instance is also used as web server and application server. Far from
ideal but has worked well for 15 months.
Problem: We have been running PostgreSQL 9.0.4 on SLES11 SP1, last kernel
in use was 2.6.32-43-0.4, performance has always been great. Since updating
from SLES11 SP1 to SP2 we now experience many database 'stalls' (e.g.
normally 'instant' queries taking many seconds, any query will be slow,
just connecting to the database will be slow). We have trialled PostgreSQL
9.2.3 under SLES11 SP2 with the exact same results. During these periods
the machine is completely responsive but anything accessing the database is
extremely slow.
I have tried increasing sched_migration_cost from 500000 to 5000000 and
also tried setting sched_compat_yield to 1, neither of these appeared to
make a difference. I don't have the parameter 'sched_autogroup_enabled'.
Nothing jumps out from top/iostat/sar/pg_stat_activity however I am very
far from expert in interpreting their output
We have work underway to reduce our number of connections as although it
has always worked ok, perhaps it makes us particularly vulnerable to
kernel/scheduler changes.
I would be very grateful for any suggestions as to the best way to
diagnose the source of this problem and/or general recommendations?

[Scott Marlowe]

There's another thread in the list started by Josh Berkus about poor
performance with a similar kernel. I'd look that thread up and see if
you and he have enough in common to work together on this.

Hardware: IBM X3650 M3 (2 x Xeon X5680 6C 3.33GHz), 96GB RAM. IBM X3524 with
RAID 10 ext4 (noatime,nodiratime,data=writeback,barrier=0) volumes for
pg_xlog / data / indexes.
Software: SLES 11 SP2 3.0.58-0.6.2-default x86_64, PostgreSQL 9.0.4.

[Mark Smith]

Post by Mark Smith
Software: SLES 11 SP2 3.0.58-0.6.2-default x86_64, PostgreSQL 9.0.4.

[skipped]

Post by Mark Smith
Problem: We have been running PostgreSQL 9.0.4 on SLES11 SP1, last

kernel in

Post by Mark Smith
use was 2.6.32-43-0.4, performance has always been great. Since updating
from SLES11 SP1 to SP2 we now experience many database 'stalls' (e.g.
normally 'instant' queries taking many seconds, any query will be slow,

just

Post by Mark Smith
connecting to the database will be slow).

It reminds me a transparent huge pages defragmentation issue that was
found in recent kernels.
Transparent huge pages defragmentation could lead to unpredictable
database stalls on some Linux kernels. The recommended settings for
this are below.
db1: ~ # echo always > /sys/kernel/mm/transparent_hugepage/enabled
db1: ~ # echo madvise > /sys/kernel/mm/transparent_hugepage/defrag
[skipped]

Sergey - your suggestion to look at transparent huge pages (THP) has
resolved the issue for us, thank you so much. We had noticed abnormally
high system CPU usage but didn't get much beyond that in our analysis.

We disabled THP altogether and it was quite simply as if we had turned the
'poor performance' tap off. Since then we have had no slow queries / stalls
at all and system CPU is consistently very low. We changed many things
whilst trying to resolve this issue but the THP change was done in
isolation and we can therefore be confident that in our environment,
leaving THP enabled with the default parameters is a killer.

At a later point we will experiment with enabling THP with the recommended
madvise defrag setting.

Thank you to all who responded.

Mark