Re: Slow nested loop execution on larger server

On Fri, Dec 16, 2011 at 1:27 PM, Greg Smith <greg(at)2ndquadrant(dot)com> wrote:
> Forwarding this on from someone who may pop up on the list too.  This is a
> new server that's running a query slowly.  The rate at which new semops
> happen may be related.  OS is standard RHEL 5.5, kernel 2.6.18-194.32.1.el5.
>  One possibly interesting part is that this is leading edge hardware:  four
> 8-core processes with HyperThreading, for 64 threads total:
>
> processor       : 63
> model name      : Intel(R) Xeon(R) CPU           X7550  @ 2.00GHz
> physical id     : 3
> siblings        : 16
> cpu cores       : 8
>
> My only vague memory of issues around systems like this that comes to mind
> here is Scott Marlowe talking about some kernel tweaks he had to do on his
> 48 core AMD boxes to get them work right.  That seems like a sketchy
> association, and I don't seem to have the details handy.
>
> I do have the full query text and associated plans, but it's mostly noise.
>  Here's the fun part:
>
> ->  Nested Loop  (cost=581.52..3930.81 rows=2 width=51) (actual
> time=25.681..17562.198 rows=21574 loops=1)
>      ->  Merge Join  (cost=581.52..783.07 rows=2 width=37) (actual
> time=25.650..98.268 rows=21574 loops=1)
>            Merge Cond: (rim.instrumentid = ri.instrumentid)
>            ->  Index Scan using reportinstruments_part_125_pkey on
> reportinstruments_part_125 rim  (cost=0.00..199.83 rows=110 width=8) (actual
> time=0.033..27.180 rows=20555 loops=1)
>                  Index Cond: (reportid = 105668)
>            ->  Sort  (cost=581.52..582.31 rows=316 width=33) (actual
> time=25.608..34.931 rows=21574 loops=1)
>                  Sort Key: ri.instrumentid
>                  Sort Method: quicksort  Memory: 2454kB
>                  ->  Index Scan using riskbreakdown_part_51_pkey on
> riskbreakdown_part_51 ri  (cost=0.00..568.40 rows=316 width=33) (actual
> time=0.019..11.599 rows=21574 loops=1)
>                        Index Cond: (reportid = 105668)
>      ->  Index Scan using energymarketlist_pkey on energymarketlist ip
>  (cost=0.00..1573.86 rows=1 width=18) (actual time=0.408..0.808 rows=1
> loops=21574)
>            Index Cond: ((reportid = 105668) AND (instrumentid =
> ri.instrumentid))
> ...
> Total runtime: 21250.377 ms
>
> The stats are terrible and the estimates off by many orders of magnitude.
>  But that's not the point.  It expected 2 rows and 21574 came out; fine.
>  Why is it taking this server 17 seconds to process 21K rows of tiny width
> through a Nested Loop?  Is it bouncing to a new CPU every time the thing
> processes a row or something?  I'm finding it hard to imagine how this could
> be a PostgreSQL problem; seems more like a kernel bug aggrevated on this
> system.  I wonder if we could produce a standalone test case with a similar
> plan from what the bad query looks like, and ask the person with this
> strange system to try it.  See if it's possible to make it misbehave in the
> same way with something simpler others can try, too.
>
> The only slow semops thread I found in the archives turned out to be I/O
> bound.  This query is all CPU; combining a few examples here since this is
> repeatable and I'm told acts the same each time:
>
>  PID USER      PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+  COMMAND
> 12634 xxx       25   0  131g 1.4g 1.4g R 100.1  0.6  11:24.20 postgres: xxx
> xxx [local] SELECT
>
> $ ./pidstat -d -T ALL -p 27683 1 100
> 15:28:46          PID   kB_rd/s   kB_wr/s kB_ccwr/s  Command
> 15:28:47        27683      0.00      0.00      0.00  postgres
>
> The PostgreSQL 9.1.1 is a custom build, but looking at pg_config I don't see
> anything special that might be related; it was optimized like this:
>
> CFLAGS=-march=core2 -O2 -pipe
>
> Here's a chunk of the observed and seemingly strange semop rate:
>
> $ strace -fttp 12634
> Process 12634 attached - interrupt to quit
> 15:38:55.966047 semop(16646247, 0x7fffd09dac20, 1) = 0
> 15:39:17.490133 semop(16810092, 0x7fffd09dac20, 1) = 0
> 15:39:17.532522 semop(16810092, 0x7fffd09dac20, 1) = 0
> 15:39:17.534874 semop(16777323, 0x7fffd09dac00, 1) = 0
> 15:39:17.603637 semop(16777323, 0x7fffd09dac00, 1) = 0
> 15:39:17.640646 semop(16810092, 0x7fffd09dac20, 1) = 0
> 15:39:17.658230 semop(16810092, 0x7fffd09dac20, 1) = 0
> 15:39:18.905137 semop(16646247, 0x7fffd09dac20, 1) = 0
> 15:39:33.396657 semop(16810092, 0x7fffd09dac20, 1) = 0
> 15:39:50.208516 semop(16777323, 0x7fffd09dac00, 1) = 0
> 15:39:54.640712 semop(16646247, 0x7fffd09dac20, 1) = 0
> 15:39:55.468458 semop(16777323, 0x7fffd09dac00, 1) = 0
> 15:39:55.488364 semop(16777323, 0x7fffd09dac00, 1) = 0
> 15:39:55.489344 semop(16777323, 0x7fffd09dac00, 1) = 0
> Process 12634 detached
>
> pg_locks for this 12634 shows all granted ones, nothing exciting there.  I
> asked how well this executes with enable_nestloop turned off, hoping to see
> that next.
>
> This all seems odd, and I get interested and concerned when that start
> showing up specifically on newer hardware.

Ridiculously late response here, but, IME, semop() calls typically
indicate LWLock contention, but with a stock build it's pretty well
impossible to figure out which LWLock is being contended; compiling
with LWLOCK_STATS could tell ou that.

Shooting from the hip, the first thing that comes to mind is that the
index isn't fully cached in shared_buffers, and every time you hit a
page that isn't there you have to acquire BufFreelistLock to run the
clock sweep. If the lock is uncontended then you wouldn't get system
calls, but if there's other activity on the system you might get
something like this.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company