Skip site navigation (1)Skip section navigation (2)
Date:      Sun, 10 Apr 2016 10:39:54 -0400
From:      PK1048 <paul@pk1048.com>
To:        Wim Lewis <wiml@omnigroup.com>
Cc:        "freebsd-fs@FreeBSD.org" <freebsd-fs@freebsd.org>
Subject:   Re: ZFS pool with a large number of filesystems
Message-ID:  <A677620F-7787-4EED-AA94-48887965F58F@pk1048.com>
In-Reply-To: <34DB45E8-7E1F-4D7C-96FF-E0A403EE8000@omnigroup.com>
References:  <34DB45E8-7E1F-4D7C-96FF-E0A403EE8000@omnigroup.com>

next in thread | previous in thread | raw e-mail | index | archive | help

> On Apr 4, 2016, at 20:38, Wim Lewis <wiml@omnigroup.com> wrote:
>=20
> I'm curious how many ZFS filesystems are reasonable to have on a =
single machine (in a single zpool). We're contemplating a design in =
which we'd have tens of thousands, perhaps a couple hundred thousand, =
filesystems mounted out of the same pool. Before we go too far into =
investigating this idea: Does anyone have real-world experience doing =
something like that? Is it a situation that ZFS-on-FreeBSD is engineered =
to handle with good performance? Is there a rough estimate of the =
resources consumed per additional filesystem (in terms of kernel VM and =
disk space)?
>=20
> Thanks for any insight or advice (even, or especially, if the answer =
is "that's crazy, don't do that" :) )

In a previous life I ran ZFS on Solaris (10) with hundreds of datasets, =
but all off of the top level zpool. We also had hundreds of thousands of =
snapshots. There was no noticeable hit to user performance, but as =
others have pointed out, getting a listing of all the datasets and =
snapshots was not a fast operation. zfs list -t filesystem on the other =
hand was fast.

Fast forwarding to today=E2=80=A6 I have a handful of FreeBSD boxes =
running ZFS with less than 100 datasets but with _hourly_ snapshots kept =
for 30 days. That is 720 snapshots per dataset. Our largest server has =
about 50 datasets, so about 36,000 snapshots. zfs list -t filesystem is =
fast while zfs list -t snapshot is slow, but I don=E2=80=99t do that =
very often. No performance hit to normal I/O.

The FreeBSD hardware is Xeon X55xx or E55xx, usually 2 quad core CPUs, =
so 8 or 16 vCPU. RAM varies from  low of 12GB to a high of 80GB. Systems =
are running a mix of vBox VMs, file service, backup service (including =
Mac OS Timemachine backups via NetATalk).





Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?A677620F-7787-4EED-AA94-48887965F58F>