|From:||"Ideriha, Takeshi" <ideriha(dot)takeshi(at)jp(dot)fujitsu(dot)com>|
|To:||"Ideriha, Takeshi" <ideriha(dot)takeshi(at)jp(dot)fujitsu(dot)com>, "Tsunakawa, Takayuki" <tsunakawa(dot)takay(at)jp(dot)fujitsu(dot)com>|
|Cc:||Andres Freund <andres(at)anarazel(dot)de>, Alvaro Herrera <alvherre(at)2ndquadrant(dot)com>, Tomas Vondra <tomas(dot)vondra(at)2ndquadrant(dot)com>, Bruce Momjian <bruce(at)momjian(dot)us>, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us>, "PostgreSQL Hackers" <pgsql-hackers(at)lists(dot)postgresql(dot)org>, Kyotaro HORIGUCHI <horiguchi(dot)kyotaro(at)lab(dot)ntt(dot)co(dot)jp>, Michael Paquier <michael(dot)paquier(at)gmail(dot)com>, 'Robert Haas' <robertmhaas(at)gmail(dot)com>, "David Steele" <david(at)pgmasters(dot)net>, Craig Ringer <craig(at)2ndquadrant(dot)com>|
|Subject:||RE: Protect syscache from bloating with negative cache entries|
|Views:||Raw Message | Whole Thread | Download mbox | Resend email|
>From: Ideriha, Takeshi [mailto:ideriha(dot)takeshi(at)jp(dot)fujitsu(dot)com]
>Does this result show that hard-limit size option with memory accounting doesn't harm
>to usual users who disable hard limit size option?
I've implemented relation cache size limitation with LRU list and built-in memory context size account.
And I'll share some results coupled with a quick recap of catcache so that we can resume discussion if needed
though relation cache bloat was also discussed in this thread but right now it's pending
and catcache feature is not fixed. But a variety of information could be good I believe.
Regarding catcache it seems to me recent Horiguchi san posts shows a pretty detailed stats
including comparison LRU overhead and full scan of hash table. According to the result, lru overhead seems small
but for simplicity this thread go without LRU.
When there was hard limit of catcach, there was built-in memory context size accounting machinery.
I checked the overhead of memory accounting, and when repeating palloc and pfree of 800 byte area many times it was 4% down
on the other hand in case of 32768 byte there seems no overhead.
Regarding relcache hard limit (relation_cache_max_size), most of the architecture was similar as catcache one with LRU list except memory accounting.
Relcaches are managed by LRU list. To prune LRU cache, we need to know overall relcache sizes including objects pointed by relcache
like 'index info'.
So in this patch relcache objects are allocated under RelCacheMemoryContext, which is child of CacheMemoryContext. Objects pointed by
relcache is allocated under child context of RelCacheMemoryContext.
In built-in size accounting, if memoryContext is set to collect "group(family) size", you can get context size including child easily.
I ran two experiments:
A) One is pgbench using Tomas's script he posted while ago, which is randomly select 1 from many tables.
B) The other is to check memory context account overhead using the same method.
A) randomly select 1 from many tables
Results are average of 5 times each.
number of tables | 100 |1000 |10000
TPS (master) |11105 |10815 |8915
TPS (patch; limit feature off) |11254 (+1%) |11176 (+3%) |9242 (+4%)
TPS (patch: limit on with 1MB) |11317 (+2%) |10491 (-3%) |7380 (-17%)
The results are noisy but it seems overhead of LRU and memory accounting is small when turning off the relcache limit feature.
When turning on the limit feature, after exceeding the limit it drops 17%, which is no surprise.
B) Repeat palloc/pfree
"With group accounting" means that account test context and its child context with built-in accounting using "palloc_bench_family()".
The other one is that using palloc_bench(). Please see palloc_bench.gz.
Master | 59.97 ms
Master with group account | 59.57 ms
patched |67.23 ms
patched with family |68.81 ms
It seems that overhead seems large in this patch. So it needs more inspection this area.
|Next Message||Michael Paquier||2019-04-11 06:57:27||Re: BUG #15672: PostgreSQL 11.1/11.2 crashed after dropping a partition table|
|Previous Message||Noah Misch||2019-04-11 06:22:28||Re: [HACKERS] Weaker shmem interlock w/o postmaster.pid|