Score:1

ZFS/ZoL: Consequences of increasing zfs_vdev_async_read_min_active (and max)?

mx flag

Vdev async read queues have min=1 and max=3 by default. All sync read/write queues are defaulted to min=max=10, while async writes get min=2 and max=10.

Async reads are described as "prefetch reads" in the docs. So I take it that by design, when i/o is heavy, prefetch activity should be kept relatively low. Yet I have seen a lot of recommended configs which raise zfs_vdev_async_read_min_active and the other min_active tunables to the same number A > 10. The corresponding maximums are typically given the same value B, and sometimes B=A.

I understand the general reason for making the i/o queues bigger, but is it a good idea to change the balance between prefetch reads and other i/o?

Perhaps, it's a good idea, only if the workload has a favorable ratio of prefetch hits to misses?

ewwhite avatar
ng flag
It depends. Are you asking about a specific situation or configuration in mind?
mx flag
Well, I've inherited a configuration where all zfs_vdev_(a)sync_(read|write)_(min|max)_active = 16. We've had some performance issues. Before diving into workload testing with different settings, I wanted to understand the theory behind this parameter.
ewwhite avatar
ng flag
What leads you to believe that the performance issues are related to the settings?
mx flag
I don't necessarily believe that, I'm just investigating.
ewwhite avatar
ng flag
There isn't enough information here to assist.
mangohost

Post an answer

Most people don’t grasp that asking a lot of questions unlocks learning and improves interpersonal bonding. In Alison’s studies, for example, though people could accurately recall how many questions had been asked in their conversations, they didn’t intuit the link between questions and liking. Across four studies, in which participants were engaged in conversations themselves or read transcripts of others’ conversations, people tended not to realize that question asking would influence—or had influenced—the level of amity between the conversationalists.