I've been struggling with transparent huge pages performance issues of late, and noticed many database systems recommend turning it off. I am talking Oracle, Postgresql, MySQL, Cassandra, NuoDB, Redis, Hadoop, and more.

For examples:

So I am wondering: what kinds of workload do benefit from this feature?

  • 2
    I'm curious as well. According to the kernel docs, they saw a big jump in kvm performance but they don't go into much detail. It's also mentioned in the kernel tuning guide for KVM.
    – Bratchley
    May 7, 2015 at 1:34
  • 4
    The main painpoint with HugePages seesm to be NUMA related. Practically every new shiny multiprocessor server out there is now NUMA enabled. Funky things in the memory subsystem ensue once you have huge pages that cross NUMA boundaries. Have a look at: engineering.linkedin.com/performance/…
    – Lmwangi
    May 23, 2015 at 20:45

2 Answers 2


Huge pages would be useful in a situation where you needed a huge amount of information to be written in the same block. It can relate to the strategy for disk writes and can be significant for caching. Like all configuration options, it makes no sense if your use case does not fit.

So the answer is, workload that actually needs a huge ton of data in the same block would benefit from huge pages. If the data is so large it would not be able to fit, but would have to be broken up into numerous page files, and the sheer number of those would be too many to handle or bad for some reason, and the lower number of larger files is preferable - there is your case for huge page files.

In practical terms, I have never come across any need for it but I know from managing caches - it is a real thing and someone, somewhere, could benefit from huge pages.

  • 3
    Yes, explicit huge pages are great for such workloads (I believe PostgreSQL uses them if available). The questioner asked about transparent huge pages, though. Jun 18, 2015 at 16:46

Don't know who told you cassandra won't benefit from hugepages. Maybe you want to be a little chatty in the direction of defrag option from /sys/kernel/mm/transparent_hugepage.

Personally just tested a cassandra cluster with and without hugepage and after various tests with different paritition size, starting from 300 b up to 4k, I can tell I am going to reenable them.

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .