[Pdns-users] Performance drop after upgrade from auth 3.4.11 to 4.0.4

[Klaus Darilion]

Hi Bert!


Am 19.09.2017 um 21:18 schrieb bert hubert:
> Can you redo your measurements against 4.1rc1? We fixed a lot in there.
> Would be interesting to know if 4.1 is already better.

Targets: reg-tst1, reg-tst2. (they are identical: 4vCPUs, 8GB RAM, no
other VMs running on this hypervisor)

Tester and Target are in the same network, 1GBit, almost idle, connected
via a Switch.

The list of testdomains has 700.000 unique entries

Testscenario: Every target host will be tested 2 times:

  *
    domU reboot (caches will be empty)
  *
    disable slony replication (no updates to the DB)
  *
    test run 1 (will be slow, all caches empty)
  *
    test run 2 (PDNS caches will be empty again as the test run lasts
    several minutes, but the Linux Filesystem Cache should be hot and
    hence DB-Queries should be much faster)

Every test will be reproduced 2 times to see how stable the results are.

For testing I am using dnsperf with 2 query-lists:

1. query the SOA for every <domain>

2. query the A for every www.<domain>

(the table is best viewed in HTML)

Test-Target 	reg-tst1 	reg-tst2
Version Server 	Run1 cold 	Run1 hot 	Run2 cold 	Run2 hot 	Run1 cold
Run1 hot 	Run2 cold 	Run2 hot
3.4.11 SOA 	2162 (lost1039) 	4167 (lost828) 	2152 (lost1183) 	4176
(lost1092) 	2171 (lost1108) 	4456 (lost953) 	2176 (lost1180) 	4459
(lost755)
3.4.11 A 	1361 (lost652) 	2463 (lost201) 	1385 (lost608) 	2432 (lost113)
	1378 (lost657) 	2568 (lost229) 	n/a
	n/a
4.0.4 SOA 	2260 (lost134) 	4546 (lost274) 	2240 (lost201) 	4625
(lost133) 	2020 (lost43) 	4900 (lost152) 	2248 (lost98) 	4900 (lost130)
4.0.4 A 	1142 (lost140) 	1781 (lost123) 	1143 (lost497) 	1783 (lost147)
1013 (lost513) 	1504 (lost101) 	1021 (lost516) 	1543 (lost132)
0.0.g916d19c SOA 	2359 (lost629) 	6141 (lost521) 	2375 (lost686) 	6152
(lost472) 	2446 (lost206) 	6779 (lost229) 	2463 (lost192) 	6795 (lost181)
0.0.g916d19c A 	1479 (lost563) 	2791 (lost1164) 	1426 (lost917) 	2763
(lost1196) 	1327 (lost689) 	2491 (lost989) 	1342 (lost596) 	2489 (lost996)

Hence, master is faster than 4.0. 4.0 is faster or slower then 3.4.11.

I suspect that the prepared-statements in 4.0 caused the performance
drop, as every DB query are now 2 DB interactions: PREPARE SELECT ... +
EXECUTE

TODO: testing with real traffic patterns


regards

Klaus



-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.powerdns.com/pipermail/pdns-users/attachments/20170921/b5b4e2da/attachment.html>