> I've tried to re-run the test for some specific values of
> effective_io_concurrency. The results were the same.
> That's why I don't think the order of tests or variability in "hardware"
> performance affected the results.
We run many MS SQL server VMs in AWS with more than adequate performance.
AWS EBS performance is variable and depends on various factors, mainly the size
of the volume and the size of the VM it is attached to. The bigger the VM, the
more EBS “bandwidth” is available, especially if the VM is EBS Optimised.
The size of the disk determines the IOPS available, with smaller disks
naturally getting less. However, even a small disk with (say) 300 IOPS is
allowed to burst up to 3000 IOPS for a while and then gets clobbered. If you
want predictable performance then get a bigger disk! If you really want
maximum, predictable performance get an EBS Optimised VM and use Provisioned
IOPS EBS volumes…. At a price!
On 31/01/2018 15:01, Rick Otten wrote:
We moved our stuff out of AWS a little over a year ago because the performance
was crazy inconsistent and unpredictable. I think they do a lot of
oversubscribing so you get strange sawtooth performance patterns depending on
who else is sharing your infrastructure and what they are doing at the time.
The same unit of work would take 20 minutes each for several hours, and then
take 2 1/2 hours each for a day, and then back to 20 minutes, and sometimes
anywhere in between for hours or days at a stretch. I could never tell the
business when the processing would be done, which made it hard for them to set
expectations with customers, promise deliverables, or manage the business.
Smaller nodes seemed to be worse than larger nodes, I only have theories as to
why. I never got good support from AWS to help me figure out what was
My first thought is to run the same test on different days of the week and
different times of day to see if the numbers change radically. Maybe spin up a
node in another data center and availability zone and try the test there too.
My real suggestion is to move to Google Cloud or Rackspace or Digital Ocean or
somewhere other than AWS. (We moved to Google Cloud and have been very happy
there. The performance is much more consistent, the management UI is more
intuitive, AND the cost for equivalent infrastructure is lower too.)
On Wed, Jan 31, 2018 at 7:03 AM, Vitaliy Garnashevich <vgarnashev...@gmail.com
<mailto:vgarnashev...@gmail.com> > wrote:
I've tried to run a benchmark, similar to this one:
CREATE TABLESPACE test OWNER postgres LOCATION '/path/to/ebs';
pgbench -i -s 1000 --tablespace=test pgbench
echo "" >test.txt
for i in 0 1 2 4 8 16 32 64 128 256 ; do
sync; echo 3 > /proc/sys/vm/drop_caches; service postgresql restart
echo "effective_io_concurrency=$i" >>test.txt
psql pgbench -c "set effective_io_concurrency=$i; set enable_indexscan=off;
explain (analyze, buffers) select * from pgbench_accounts where aid between
1000 and 10000000 and abalance != 0;" >>test.txt
I get the following results:
Execution time: 40262.781 ms
Execution time: 98125.987 ms
Execution time: 55343.776 ms
Execution time: 52505.638 ms
Execution time: 54954.024 ms
Execution time: 54346.455 ms
Execution time: 55196.626 ms
Execution time: 55057.956 ms
Execution time: 54963.510 ms
Execution time: 54339.258 ms
The test was using 100 GB gp2 SSD EBS. More detailed query plans are attached.
PostgreSQL 9.6.6 on x86_64-pc-linux-gnu, compiled by gcc (Ubuntu
5.4.0-6ubuntu1~16.04.4) 5.4.0 20160609, 64-bit
The results look really confusing to me in two ways. The first one is that I've
seen recommendations to set effective_io_concurrency=256 (or more) on EBS. The
other one is that effective_io_concurrency=1 (the worst case) is actually the
default for PostgreSQL on Linux.