On 5/7/21 9:06 PM, Yura Sokolov wrote:
Pavel Stehule писал 2021-05-07 21:45:
Samples: 119K of event 'cycles', 4000 Hz, Event count (approx.):
Overhead Shared Object Symbol
79.20% postgres [.] cache_reduce_memory
1.94% [kernel] [k] native_write_msr_safe
1.63% [kernel] [k] update_cfs_shares
1.00% [kernel] [k] trigger_load_balance
0.97% [kernel] [k] timerqueue_add
0.51% [kernel] [k] task_tick_fair
0.51% [kernel] [k] task_cputime
0.50% [kernel] [k] perf_event_task_tick
0.50% [kernel] [k] update_curr
0.49% [kernel] [k] hrtimer_active
Regards
Pavel
It is strange to see cache_reduce_memory itself consumes a lot of CPU.
It doesn't contain CPU hungry code.
It calls prepare_probe_slot, that calls some tuple forming. Then
it calls resultcache_lookup that may call to ResultCacheHash_hash
and ResultCacheHash_equal. And finally it calls remove_cache_entry.
I suppose remove_cache_entry should consume most of CPU time since
it does deallocations.
And if you compile with --enable-cassert, then remove_cache_entry
iterates through whole cache hashtable, therefore it reaches
quadratic complexity easily (or more correct M*N, where M - size
of a table, N - eviction count).
Yeah. I tried reproducing the issue, but without success ...
Not sure what's wrong, but --enable-cassert is one option. Or maybe
there's some funny behavior due to collecting timing info?
FWIW the timings on my laptop look like this:
work_mem=40MB 5065ms
work_mem=10MB 5104ms
resultcache=off 13453ms
So a very different behavior from what Pavel reported. But if I rebuild
with casserts, I get the same massive slowdown, so I guess that's it.
regards
--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company