Jul 18 2021 06:22 PM
Jul 18 2021 06:22 PM
A simple issue demonstration script below.
On some servers (mostly development, physical and VMs with 2-4 cores) it is executed ~3 seconds.
On other servers (mostly production VMs with 16-20 cores) it is executed 18-22 seconds.
Mind-blowing ~5 times slower.
A PowerShell script doing the same work shows reasonable similar execution time on all those servers.
"Mostly", but there are exceptions killing any idea we came up with.
i.e. "Enterprise and Development editions are different" - have 2 fast Enterprise servers.
i.e. "Many cores is not good" - have a fast server with 4 cores and slow server with 4 cores.
i.e. "Your physical server may have different CPU" - have a fast and slow VMs on the same ESXi box.
i.e. "Your physical server is very busy" - have a slow (not currently loaded) VM on dedicated ESXi box.
Any idea how to troubleshot this?
Most wild observation - the usually slow server is fast during first minute after reboot.
Like another delayed-start service creates the problem. But I cannot imagine the logic of such interference ...
declare @start datetime = getdate()
declare @sum bigint = 0
declare @i int = 0
while @i < 4000000
set @sum += @i
set @i += 1
print datediff(millisecond, @start, getdate())
Jul 19 2021 12:38 PM
Hi @AnatoliD -- We will probably need a bit more information in order to assist.
Jul 19 2021 09:52 PM - edited Jul 19 2021 10:02 PM
1. Majority are 2017, but noticed on 2019 and 2014. Yes. Majority of slow servers are prod, but I have one slow and one fast both Prod (Enterprise edition) both running on the same physical ESXi.
2. This code does not touch any table and SSMS/SqlSentry does not return actual execution plan. Estimated execution plan is the same (COND + ASSIGNs)
3. Default 0 everywhere. The loop is single thread, it cannot be parallelized anyway.
Jul 21 2021 07:01 PM
Hi @AnatoliD -- I spent some time working with this on various platforms (virtualized, PaaS) with various versions of SQL Server running on both Windows and Linux. I'm not really seeing any consistency or patterns with execution times. The vast majority of the time the loop executes in less than three seconds. Is this related to an application/database issue or is this more of a proof-of-concept? Take care.
Jul 21 2021 10:25 PM
HI@bake13 Thanks for looking into this.
I have performance issues with a few "applications" (a set of SPs/UDFs for data processing / transformation). It runs fine on dev with 2 cores, and not fine on prod with 16 cores.
I don't really hope it would work 8 times faster on prod, but 6 times slower - totally unacceptable.
This test is free of any other factors (disk/network IO, blocking, old statistics and not optimal execution plan, etc) which could affect performance of complex system.