Re: Distributing data over "spindles" even on AWS EBS, (followup to the work queue saga) - Mailing list pgsql-performance
I am going to reply to my own message to add more information that I think can be interesting for others.
I find that IO contention continues being a number one issue for streamlining database performance (it's a duh! but it's significant sticky point!)
I told you my performance was again going down the drain rapidly. In my case it's going so fast because the database is filling up fast with all that heavy load activity. I still don't know if autovacuum is running well. But when I tried to manually vacuumdb I noticed the number of dead rows relatively small, so I don't think that the lack of vacuuming was an issue.
I have a bottleneck though.
There is one massive table, let's call it Foo, and in Foo there is also significant amount of toasted text, and there is a child table called Foo_id then two indexes on Foo_id. Here the essentials:
CREATE TABLE Foo { internalId UUID PRIMARY KEY, ... -- tons of columns text_xml text, -- lot's of stuff to toast ... -- tons of more columns
};
CREATE TABLE Foo_id { fooInternalId UUID REFERENCES Foo(internalId), prefix text, suffix text
}
CREATE INDEX Foo_id_fkidx ON Foo_id(fooInternalId);
CREATE INDEX Foo_id_idx ON Foo_id(prefix, suffix);
Now, it so happens that the activity on that index is so large that this volume has 100% io utilization per iostat and everyone is at 75% iowait:Device: rrqm/s wrqm/s r/s w/s rkB/s wkB/s avgrq-sz avgqu-sz await r_await w_await svctm %util nvme8n1 0.00 0.00 69.00 45.00 880.00 400.00 22.46 8.79 90.95 88.64 94.49 8.77 100.00
However, the only really heavily used data file here really is just that Foo_id_idx. That one index!
I think I have a few ways to alleviate the bottleneck:
- partition that Foo_id table so that that index would be partitioned too and I can spread it out over multiple volumes.
- build a low level "spreading" scheme which is to take the partial files 4653828 and 4653828.1, .2, _fsm, etc. and move each to another device and then symlink it back to that directory (I come back to this!)
- maybe I can only partition the index by using the WHERE clause on the CREATE INDEX over some hash function of the Foo_id.prefix column.
- maybe I can configure in AWS EBS to reserve more IOPS -- but why would I pay for more IOPS if my cost is by volume size? I can just make another volume? or does AWS play a similar trick on us with IOPS being limited on some "credit" system???
To 1. I find it surprisingly complicated to have to create a partitioned table only in order to spread an index over multiple volumes.
To 2. I find that it would be a nice feature of PostgreSQL if we could just use symlinks and a symlink rule, for example, when PostgreSQL finds that 4653828 is in fact a symlink to /otherdisk/PG/16284/4653828, then it would
- by default also create 4653828.1 as a symlink and place the actual data file on /otherdisk/PG/16284/4653828.1
- or even easier: it would allow the admin to pre-create the datafiles or even just the symlinks to not yet extisting datafiles, so that when it comes to create the 4653828.2 it will do it wherever that symlink points to, either a broken symlink whose target will then be created, or a symlink to a zero-size file that was already pre-created.
- and also easier, allow us to pre-set the size of the data files with truncate --size $target_size so that PostgreSQL will use the next .1, .2, .3 file once the target size has been reached, rather than filling up all those 4 GB
- I think that if, as I find, the wisdom to "divide data over spindles" is still true, then it would be best if PostgreSQL had a distribution scheme which is not at the logical data model level (partition ... tablespace), but rather just on the low level.
That last point I just made up. But it is extremely useful if PostgreSQL would have this sort of very very simple intelligence.
To 3. if I create the index like this:
CREATE INDEX Foo_id_idx ON Foo_id(prefix, suffix) WHERE hashmod(prefix,4) = 0 TABLESPACE /tbs/Foo_id_0; CREATE INDEX Foo_id_idx ON Foo_id(prefix, suffix) WHERE hashmod(prefix,4) = 1 TABLESPACE /tbs/Foo_id_1; CREATE INDEX Foo_id_idx ON Foo_id(prefix, suffix) WHERE hashmod(prefix,4) = 2 TABLESPACE /tbs/Foo_id_2; CREATE INDEX Foo_id_idx ON Foo_id(prefix, suffix) WHERE hashmod(prefix,4) = 3 TABLESPACE /tbs/Foo_id_3;
with some appropriately defined hashmod function that divides up 4 approximately equal partitions.
Is there any downside to this approach? It looks to me that this does everything partitioning scheme would also do, i.e., (1) routing inserted tuples into the right file, and (2) resolving which file to refer to based on the data of the query. What am I missing?
regards,
-Gunther
Hello again. You may remember my queue issue for which some of you have proposed to use a partitioned table approach. I have done that, and I might report more on that once I have this beast all tamed, which may be now. Let's say in short, it definitely helped immensely. My test case now is different from what I had previously done. I am now hammering my database with 52 worker threads uploading like crazy into some 100 tables and indexes.
Right now I want to remind everybody of the surprising fact that the old wisdom of distributing load over "spindles" appears to be still true even in the virtualized world of cloud computing. For background, this is running on Amazon AWS, the db server is a c5.xlarge and you see I have 0.0 st, because my virtual CPUs are dedicated.
I had run into a situation which was totally crazy. Here I show you a snapshot of top and iostat output as it ran all night with totally low tps.
top - 12:43:42 up 1 day, 9:29, 3 users, load average: 41.03, 39.58, 38.91 Tasks: 385 total, 1 running, 169 sleeping, 0 stopped, 0 zombie %Cpu(s): 2.9 us, 0.9 sy, 0.0 ni, 5.9 id, 90.3 wa, 0.0 hi, 0.1 si, 0.0 st KiB Mem : 7809760 total, 130528 free, 948504 used, 6730728 buff/cache KiB Swap: 0 total, 0 free, 0 used. 4357496 avail Mem PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 20839 postgres 20 0 2309448 86892 83132 D 1.0 1.1 0:00.04 postgres: auser integrator 172.31.49.159(44862) SELECT 17230 postgres 20 0 2318736 1.7g 1.7g D 0.7 23.0 0:10.00 postgres: auser integrator 172.31.49.159(44458) SELECT 19209 postgres 20 0 2318760 1.7g 1.7g D 0.7 22.7 0:04.89 postgres: auser integrator 172.31.54.158(44421) SELECT 19467 postgres 20 0 2318160 1.8g 1.8g D 0.7 23.6 0:04.20 postgres: auser integrator 172.31.61.242(56981) INSERT 19990 postgres 20 0 2318084 1.2g 1.2g D 0.7 16.1 0:02.12 postgres: auser integrator 172.31.63.71(50413) SELECT 20004 postgres 20 0 2317924 863460 853052 D 0.7 11.1 0:02.10 postgres: auser integrator 172.31.63.71(21895) INSERT 20555 postgres 20 0 2316952 899376 890260 D 0.7 11.5 0:00.65 postgres: auser integrator 172.31.61.242(60209) INSERT 20786 postgres 20 0 2312208 736224 729528 D 0.7 9.4 0:00.22 postgres: auser integrator 172.31.63.71(48175) INSERT 18709 postgres 20 0 2318780 1.9g 1.8g D 0.3 24.9 0:06.18 postgres: auser integrator 172.31.54.158(17281) SELECT 19228 postgres 20 0 2318940 1.7g 1.7g D 0.3 22.4 0:04.63 postgres: auser integrator 172.31.63.71(63850) INSERT 19457 postgres 20 0 2318028 1.1g 1.1g D 0.3 15.0 0:03.69 postgres: auser integrator 172.31.54.158(33298) INSERT 19656 postgres 20 0 2318080 1.3g 1.3g D 0.3 18.1 0:02.90 postgres: auser integrator 172.31.61.242(23307) INSERT 19723 postgres 20 0 2317948 1.3g 1.2g D 0.3 16.8 0:02.17 postgres: auser integrator 172.31.49.159(44744) SELECT 20034 postgres 20 0 2318044 927200 916924 D 0.3 11.9 0:02.19 postgres: auser integrator 172.31.63.71(64385) SELECT 20080 postgres 20 0 2318124 1.2g 1.2g D 0.3 15.6 0:01.90 postgres: auser integrator 172.31.63.71(23430) INSERT 20264 postgres 20 0 2317824 1.0g 1.0g D 0.3 13.9 0:01.28 postgres: auser integrator 172.31.54.158(64347) INSERT 20285 postgres 20 0 2318096 582712 572456 D 0.3 7.5 0:01.08 postgres: auser integrator 172.31.63.71(34511) INSERT 20392 root 20 0 0 0 0 I 0.3 0.0 0:00.05 [kworker/u8:1] 19954 postgres 20 0 2317848 1.2g 1.2g D 0.3 15.8 0:01.95 postgres: auser integrator 172.31.61.242(65080) SELECT 20004 postgres 20 0 2317924 863460 853052 D 0.3 11.1 0:02.08 postgres: auser integrator 172.31.63.71(21895) INSERT 20034 postgres 20 0 2318044 923876 913600 D 0.3 11.8 0:02.18 postgres: auser integrator 172.31.63.71(64385) SELECT 20080 postgres 20 0 2318124 1.2g 1.1g D 0.3 15.6 0:01.89 postgres: auser integrator 172.31.63.71(23430) SELECT 20248 postgres 20 0 2318312 598416 587972 D 0.3 7.7 0:01.14 postgres: auser integrator 172.31.63.71(44375) SELECT 20264 postgres 20 0 2317824 1.0g 1.0g D 0.3 13.9 0:01.27 postgres: auser integrator 172.31.54.158(64347) INSERT 20350 postgres 20 0 2318228 546652 536396 D 0.3 7.0 0:00.87 postgres: auser integrator 172.31.54.158(60787) INSERT 20590 postgres 20 0 2317208 893232 883840 D 0.3 11.4 0:00.61 postgres: auser integrator 172.31.61.242(14003) INSERT 20595 postgres 20 0 2317172 884792 875428 D 0.3 11.3 0:00.59 postgres: auser integrator 172.31.54.158(59843) INSERT 20603 postgres 20 0 2316596 838408 829668 D 0.3 10.7 0:00.50 postgres: auser integrator 172.31.61.242(16697) INSERT 20770 postgres 20 0 171388 4456 3628 R 0.3 0.1 0:00.13 top -cyou can see here that all these postgress processes are in "non-interruptible sleep" (D) state. CPU% is ridiculously low (and that's not because of steal, c5 instances do not run on "CPU credits"). Are they all in IO blocked state? Let's see iostat:
avg-cpu: %user %nice %system %iowait %steal %idle 2.51 0.00 0.75 94.99 0.00 1.75 Device: rrqm/s wrqm/s r/s w/s rkB/s wkB/s avgrq-sz avgqu-sz await r_await w_await svctm %util nvme1n1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 nvme2n1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 nvme3n1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 nvme4n1 0.00 0.00 1.00 5.00 8.00 27.50 11.83 0.00 0.00 0.00 0.00 0.00 0.00 nvme8n1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 nvme9n1 0.00 0.00 91.00 2.00 3040.00 3.00 65.44 0.00 0.65 0.66 0.00 0.04 0.40 nvme11n1 0.00 2.00 0.00 24.00 0.00 1090.00 90.83 0.00 0.00 0.00 0.00 0.00 0.00 nvme10n1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 nvme6n1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 nvme7n1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 nvme12n1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 nvme5n1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 nvme16n1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 nvme15n1 0.00 0.00 0.00 2.00 0.00 1.50 1.50 0.00 0.00 0.00 0.00 0.00 0.00 nvme13n1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 nvme14n1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 nvme17n1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 nvme18n1 0.00 0.00 194.00 133.00 1896.00 3253.50 31.50 6.90 23.27 31.18 11.73 2.46 80.40 nvme19n1 0.00 0.00 6.00 13.00 48.00 355.50 42.47 0.00 0.00 0.00 0.00 0.00 0.00 nvme20n1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 nvme21n1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 nvme22n1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 nvme23n1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 nvme0n1 0.00 0.00 0.00 7.00 0.00 69.50 19.86 0.00 0.00 0.00 0.00 0.00 0.00You see that I already did a lot to balance IO out to many different tablespaces that's why there are so many volumes. Yet my iowait % was at > 95%. I though all my user data was spread out over the tablespaces, so that I could control the IO contention. But there remained a crazy hotspot on this nvme18n1 volume. And it turns out that was the data/base default tablespace, and that I had failed to actually assign proper tablespaces to many of the tables.
Now I brought the server to a maintenance halt killing and blocking all the worker threads from connecting again during the move, and then did the tablespace move.
ALTER DATABASE integrator CONNECTION LIMIT 0; SELECT pg_terminate_backend(pid) FROM pg_stat_activity WHERE datname = 'integrator' AND pid <> pg_backend_pid() AND backend_type = 'client backend'; ALTER TABLE integrator.... SET TABLESPACE ...; ... ALTER TABLE integrator.... SET TABLESPACE ...; ALTER DATABASE integrator CONNECTION LIMIT -1;And then look at what this helped:
avg-cpu: %user %nice %system %iowait %steal %idle 38.90 0.00 10.47 13.97 0.00 36.66 Device: rrqm/s wrqm/s r/s w/s rkB/s wkB/s avgrq-sz avgqu-sz await r_await w_await svctm %util nvme1n1 0.00 0.00 9.00 37.00 72.00 296.00 16.00 0.00 0.52 0.44 0.54 0.00 0.00 nvme2n1 0.00 0.00 129.00 467.00 1152.00 4152.00 17.80 0.13 0.47 0.25 0.53 0.18 10.80 nvme3n1 0.00 0.00 8.00 38.00 64.00 304.00 16.00 0.00 0.61 0.50 0.63 0.00 0.00 nvme4n1 0.00 0.00 8.00 43.00 64.00 344.00 16.00 0.00 0.47 0.00 0.56 0.00 0.00 nvme8n1 0.00 0.00 326.00 1104.00 3452.00 10248.00 19.16 0.56 0.58 0.39 0.64 0.15 20.80 nvme9n1 0.00 0.00 29.00 71.00 232.00 568.00 16.00 0.00 0.64 0.41 0.73 0.00 0.00 nvme11n1 0.00 0.00 0.00 193.00 0.00 37720.00 390.88 0.66 4.15 0.00 4.15 0.58 11.20 nvme10n1 0.00 0.00 185.00 281.00 1560.00 2264.00 16.41 0.06 0.51 0.58 0.46 0.10 4.80 nvme6n1 0.00 0.00 14.00 137.00 112.00 1096.00 16.00 0.00 0.42 0.00 0.47 0.03 0.40 nvme7n1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 nvme12n1 0.00 0.00 267.00 584.00 2656.00 4864.00 17.67 0.23 0.53 0.54 0.53 0.19 16.40 nvme5n1 0.00 0.00 22.00 14.00 176.00 112.00 16.00 0.00 0.78 1.09 0.29 0.00 0.00 nvme16n1 0.00 0.00 75.00 179.00 732.00 1432.00 17.04 0.01 0.55 0.32 0.65 0.05 1.20 nvme15n1 0.00 0.00 0.00 16.00 0.00 128.00 16.00 0.00 1.25 0.00 1.25 0.00 0.00 nvme13n1 0.00 0.00 185.00 631.00 1804.00 5904.00 18.89 0.21 0.47 0.28 0.53 0.16 13.20 nvme14n1 0.00 0.00 141.00 227.00 1128.00 1816.00 16.00 0.02 0.48 0.57 0.42 0.05 2.00 nvme17n1 0.00 0.00 69.00 250.00 704.00 2000.00 16.95 0.00 0.44 0.41 0.45 0.00 0.00 nvme18n1 0.00 0.00 9.00 9.00 72.00 72.00 16.00 0.00 0.00 0.00 0.00 0.00 0.00 nvme19n1 0.00 0.00 137.00 294.00 1576.00 3088.00 21.64 0.07 0.56 0.82 0.44 0.14 6.00 nvme20n1 0.00 0.00 191.00 693.00 1796.00 6336.00 18.40 0.37 0.65 0.44 0.70 0.20 18.00 nvme21n1 0.00 0.00 90.00 140.00 856.00 1120.00 17.18 0.01 0.56 0.36 0.69 0.05 1.20 nvme22n1 0.00 0.00 426.00 859.00 4016.00 7272.00 17.57 0.40 0.54 0.60 0.52 0.14 18.40 nvme23n1 0.00 0.00 512.00 916.00 5076.00 10288.00 21.52 0.50 0.53 0.36 0.63 0.12 17.20 nvme0n1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00And top:
top - 18:08:13 up 1 day, 14:54, 10 users, load average: 4.89, 6.09, 4.93 Tasks: 395 total, 4 running, 161 sleeping, 0 stopped, 0 zombie %Cpu(s): 55.6 us, 8.8 sy, 0.0 ni, 18.9 id, 14.2 wa, 0.0 hi, 2.4 si, 0.0 st KiB Mem : 7809760 total, 136320 free, 610204 used, 7063236 buff/cache KiB Swap: 0 total, 0 free, 0 used. 4693632 avail Mem PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 13601 postgres 20 0 2319104 1.9g 1.9g S 40.2 25.9 0:18.76 postgres: auser integrator 172.31.54.158(15235) idle 13606 postgres 20 0 2318832 1.6g 1.6g S 18.6 21.7 0:14.25 postgres: auser integrator 172.31.54.158(49226) idle i+ 13760 postgres 20 0 2318772 1.7g 1.7g S 17.6 23.4 0:11.09 postgres: auser integrator 172.31.57.147(45312) idle i+ 13600 postgres 20 0 2318892 1.9g 1.9g R 15.6 26.1 0:20.08 postgres: auser integrator 172.31.54.158(63958) BIND 13603 postgres 20 0 2318480 1.8g 1.8g S 15.3 24.0 0:22.72 postgres: auser integrator 172.31.57.147(23817) idle i+ 13714 postgres 20 0 2318640 1.8g 1.8g S 15.3 24.0 0:10.99 postgres: auser integrator 172.31.63.71(58893) idle in+ 13607 postgres 20 0 2318748 1.9g 1.9g S 14.6 25.8 0:19.59 postgres: auser integrator 172.31.57.147(11889) idle 13844 postgres 20 0 2318260 730388 719972 S 13.0 9.4 0:02.03 postgres: auser integrator 172.31.61.242(58949) idle i+ 13716 postgres 20 0 2318816 1.8g 1.8g S 12.3 24.2 0:11.94 postgres: auser integrator 172.31.63.71(53131) idle in+ 13717 postgres 20 0 2318752 1.6g 1.6g S 10.3 21.0 0:13.39 postgres: auser integrator 172.31.63.71(19934) idle in+ 13837 postgres 20 0 2318296 805832 795380 S 10.3 10.3 0:02.28 postgres: auser integrator 172.31.61.242(63185) idle i+ 13839 postgres 20 0 2317956 722788 712532 S 10.3 9.3 0:02.04 postgres: auser integrator 172.31.49.159(57414) idle i+ 13836 postgres 20 0 2318188 697716 687224 R 10.0 8.9 0:02.09 postgres: auser integrator 172.31.61.242(51576) INSERT 13846 postgres 20 0 2317716 1.3g 1.3g S 10.0 17.0 0:02.19 postgres: auser integrator 172.31.61.242(16349) idle i+ 13854 postgres 20 0 2313504 224276 216592 S 7.3 2.9 0:00.42 postgres: auser integrator 172.31.61.242(18867) idle i+ 18055 postgres 20 0 2308060 2.1g 2.1g S 7.0 27.6 3:04.07 postgres: checkpointer 13602 postgres 20 0 2319160 1.8g 1.8g S 6.6 23.9 0:21.21 postgres: auser integrator 172.31.54.158(45183) idle i+ 13833 postgres 20 0 2317848 879168 869312 S 6.0 11.3 0:02.90 postgres: auser integrator 172.31.61.242(47892) idle i+ 13710 postgres 20 0 2318856 1.4g 1.4g S 5.6 19.3 0:09.89 postgres: auser integrator 172.31.63.71(22184) idle in+ 13809 postgres 20 0 2318168 1.1g 1.1g D 4.7 14.4 0:04.94 postgres: auser integrator 172.31.63.71(44808) SELECT 13843 postgres 20 0 2318276 595844 585432 S 4.0 7.6 0:01.36 postgres: auser integrator 172.31.61.242(39820) idle i+ 13860 postgres 20 0 2311872 139372 133356 R 3.7 1.8 0:00.11 postgres: auser integrator 172.31.49.159(57420) idle i+ 462 root 20 0 0 0 0 S 1.7 0.0 1:41.44 [kswapd0] 13859 postgres 20 0 2308104 96788 93884 S 1.7 1.2 0:00.05 postgres: auser integrator 172.31.61.242(43391) idle i+ 18057 postgres 20 0 2305108 19108 18624 S 1.7 0.2 1:26.62 postgres: walwriter 1559 root 0 -20 0 0 0 I 0.3 0.0 0:19.21 [kworker/1:1H] 1560 root 0 -20 0 0 0 I 0.3 0.0 0:25.19 [kworker/3:1H] 2619 root 20 0 13144 400 292 S 0.3 0.0 0:28.52 /sbin/rngd -fThis helped!
There is a nice saturation now of CPU at the high end of the "linear" range, i.e., we aren't in the distortion range or >90% and yet all workers are runn
I can do 17 transactions per second with 52 parallel worker threads. Now we have not run yet for over an hour but so far so good. With my old non-partitioned work queue table I would have long run into the index degradation.
I am not sure my autovacuum setup is working right though. I wonder if there isn't some autovacuum statistics which I can query that would give me confidence that it is actually running?
Finally the last question for now: I would like to set the XFS (all file systems are XFS) block size to the same size as the PostgreSQL page size. I am surprized this isn't a recommended action to take? It would seem to make sense to reduce IO system calls and push entire pages in one fell swoop every time. Right?
regards,
-GuntherPS: aaaaand we're going down. Ran vacuumdb again, but that didn't help much. It's going down again.
Attachment
pgsql-performance by date: