Comments on: Settling the Myth of Transparent HugePages for Databases https://www.percona.com/blog/settling-the-myth-of-transparent-hugepages-for-databases/ Mon, 23 Dec 2019 15:50:04 +0000 hourly 1 https://wordpress.org/?v=6.5.2 By: ibrar https://www.percona.com/blog/settling-the-myth-of-transparent-hugepages-for-databases/#comment-10971574 Mon, 23 Dec 2019 15:50:04 +0000 https://www.percona.com/blog/?p=55253#comment-10971574 Thanks for the report @Nan, fixed.

]]> By: Nan Xiao https://www.percona.com/blog/settling-the-myth-of-transparent-hugepages-for-databases/#comment-10971570 Mon, 23 Dec 2019 06:57:06 +0000 https://www.percona.com/blog/?p=55253#comment-10971570 TLB is Translation Lookaside Buffer, not Table Lookaside buffer.

]]> By: Alicja Kucharczyk https://www.percona.com/blog/settling-the-myth-of-transparent-hugepages-for-databases/#comment-10970432 Thu, 07 Mar 2019 21:24:33 +0000 https://www.percona.com/blog/?p=55253#comment-10970432 In my tests THP perfomed very well: https://www.postgresql.eu/events/pgconfeu2018/sessions/session/2125/slides/103/VMandFS_tuning.pdf
(page 35). My tests were different, made on smaller machine and not with pgbench, but production data by replaying the real workload and yes, surprisingly THP gave better results. Although I’m still not sure if THP are safe in production environment with database that’s running many months or years without restarting. Primarily because of memory fragmentation and spikes in CPU usage which could interfere with processing of constant t. huge pages creation and eases.

]]> By: Ibrar Ahmed https://www.percona.com/blog/settling-the-myth-of-transparent-hugepages-for-databases/#comment-10970430 Thu, 07 Mar 2019 17:24:59 +0000 https://www.percona.com/blog/?p=55253#comment-10970430 @Fernando I will share steps with you. I am checking the availability of same machine to debug the reasoning. I got a bit different results on another machine.

]]> By: Fernando Laudares Camargos https://www.percona.com/blog/settling-the-myth-of-transparent-hugepages-for-databases/#comment-10970429 Thu, 07 Mar 2019 14:36:14 +0000 https://www.percona.com/blog/?p=55253#comment-10970429 JFG: I believe Ibrar was referring to his previous blog post, https://www.percona.com/blog/2018/12/20/benchmark-postgresql-with-linux-hugepages/

Ibrar: could you please provide exact steps to reproduce your experiments with “static” hugepages for the above blog post? I for one wasn’t able to get close to such optimistic performance with hugepages, though I’ve used different parameters in my own tests: https://fosdem.org/2019/schedule/event/hugepages_databases/
My tests with THP there were only superficial though, and I didn’t paid attention to use a recent enough kernel for that.

]]> By: Jean-Christophe https://www.percona.com/blog/settling-the-myth-of-transparent-hugepages-for-databases/#comment-10970428 Thu, 07 Mar 2019 13:44:46 +0000 https://www.percona.com/blog/?p=55253#comment-10970428 Thank you for the clarification!

]]> By: Ibrar Ahmed https://www.percona.com/blog/settling-the-myth-of-transparent-hugepages-for-databases/#comment-10970426 Thu, 07 Mar 2019 12:12:49 +0000 https://www.percona.com/blog/?p=55253#comment-10970426 @Jean-Christophe check any “nmon” file you will see kernel version at bottom.

https://github.com/Percona-Lab-results/THP-POSTGRESQL-2019/tree/master/results/pgbench-tpcb/10m/disabled-thp

]]> By: Ibrar Ahmed https://www.percona.com/blog/settling-the-myth-of-transparent-hugepages-for-databases/#comment-10970425 Thu, 07 Mar 2019 12:05:39 +0000 https://www.percona.com/blog/?p=55253#comment-10970425 Sorry, I double checked and it is

$ uname -a
Linux smblade01 4.15.0-42-generic #45~16.04.1-Ubuntu SMP Mon Nov 19 13:02:27 UTC 2018 x86_64 x86_64 x86_64 GNU/Linux

]]> By: Jean-Christophe https://www.percona.com/blog/settling-the-myth-of-transparent-hugepages-for-databases/#comment-10970424 Thu, 07 Mar 2019 11:49:45 +0000 https://www.percona.com/blog/?p=55253#comment-10970424 In your article you mention and I quote “Ubuntu 16.04.4, kernel 4.13.0-36-generic”. Could you confirm again the Linux kernel version used as I am now confused?

]]> By: Ibrar Ahmed https://www.percona.com/blog/settling-the-myth-of-transparent-hugepages-for-databases/#comment-10970423 Thu, 07 Mar 2019 11:24:32 +0000 https://www.percona.com/blog/?p=55253#comment-10970423 I did that on Linux kernel version 4.15.0-42-generic #45~16.04.1-Ubuntu.

]]> By: Twirrim (@Twirrim) https://www.percona.com/blog/settling-the-myth-of-transparent-hugepages-for-databases/#comment-10970420 Wed, 06 Mar 2019 22:55:35 +0000 https://www.percona.com/blog/?p=55253#comment-10970420 For what it’s worth, there have been some significant improvements made to THP implementation in the kernel over the last 18 months, particularly in the 4.14 release: https://kernelnewbies.org/Linux_4.14?#Memory_management

I know they’ve been backported in to Oracle’s UEK kernel shipped on Oracle Linux, not sure if other distributions have picked them up as well. In particular those THP changes have dealt with some nastiness affecting machines under heavy memory pressure (e.g. Kafka machines)

]]> By: Ibrar Ahmed https://www.percona.com/blog/settling-the-myth-of-transparent-hugepages-for-databases/#comment-10970419 Wed, 06 Mar 2019 20:52:49 +0000 https://www.percona.com/blog/?p=55253#comment-10970419 You mean performance improvement in THP or classic HugePages? Do you have link too?

]]> By: Ibrar Ahmed https://www.percona.com/blog/settling-the-myth-of-transparent-hugepages-for-databases/#comment-10970418 Wed, 06 Mar 2019 20:51:57 +0000 https://www.percona.com/blog/?p=55253#comment-10970418 I think I have pointed out some of THP problem in this blog. In any case complete detail nmon system repots are in github pages. I will also try to compare these too.

]]> By: Nerijus https://www.percona.com/blog/settling-the-myth-of-transparent-hugepages-for-databases/#comment-10970417 Wed, 06 Mar 2019 19:58:20 +0000 https://www.percona.com/blog/?p=55253#comment-10970417 Could you try with more recent kernel? Where was some performance fixes for huge pages

]]> By: Jean-François Gagné https://www.percona.com/blog/settling-the-myth-of-transparent-hugepages-for-databases/#comment-10970416 Wed, 06 Mar 2019 19:51:10 +0000 https://www.percona.com/blog/?p=55253#comment-10970416 Are the results about “HugePages give really good performance” published somewhere ?

Also, do you have some intuition about why THP giving no performance gains vs HP giving good gains ?

Thanks.

]]> By: Ibrar Ahmed https://www.percona.com/blog/settling-the-myth-of-transparent-hugepages-for-databases/#comment-10970415 Wed, 06 Mar 2019 18:54:10 +0000 https://www.percona.com/blog/?p=55253#comment-10970415 No I did not tried that on FreeBSD.

]]> By: Ibrar Ahmed https://www.percona.com/blog/settling-the-myth-of-transparent-hugepages-for-databases/#comment-10970414 Wed, 06 Mar 2019 18:53:23 +0000 https://www.percona.com/blog/?p=55253#comment-10970414 Bruce, I did that, HugePages give really good performance.

]]> By: Federico Razzoli https://www.percona.com/blog/settling-the-myth-of-transparent-hugepages-for-databases/#comment-10970413 Wed, 06 Mar 2019 14:48:48 +0000 https://www.percona.com/blog/?p=55253#comment-10970413 Interesting, thank you. I am curious, do you have some info to share about THP on FreeBSD?

]]> By: Bruce Momjian https://www.percona.com/blog/settling-the-myth-of-transparent-hugepages-for-databases/#comment-10970412 Wed, 06 Mar 2019 14:46:55 +0000 https://www.percona.com/blog/?p=55253#comment-10970412 Are you planning to test non-transparent huge pages soon? That would be good.

]]>