Skip site navigation (1)Skip section navigation (2)
Date:      Sun, 13 Aug 2017 07:13:20 -0400
From:      Tenzin Lhakhang <tenzin.lhakhang@gmail.com>
To:        "Eugene M. Zheganin" <emz@norma.perm.ru>
Cc:        "freebsd-fs@FreeBSD.org" <freebsd-fs@freebsd.org>, freebsd-stable <freebsd-stable@freebsd.org>
Subject:   Re: zfs listing and CPU
Message-ID:  <CALcn87zig7b37Q_JByVMrWyPPF=qr%2BWO%2BcB4aP4-wFp-_B4FOw@mail.gmail.com>
In-Reply-To: <399c4309-c7a6-a7e9-299f-9675224c090d@norma.perm.ru>
References:  <aa26e888-05ef-c876-abf3-778ff08f4857@norma.perm.ru> <E3238E24-3AFA-4F2C-A299-D52E4D152097@kraus-haus.org> <399c4309-c7a6-a7e9-299f-9675224c090d@norma.perm.ru>

next in thread | previous in thread | raw e-mail | index | archive | help
You may want to have an async zfs-get program/script that regularly does a
zfs get -Ho and stores then in a local cache (redis or your own program) at
a set interval and then the api can hit the cache instead of directly
running get or list.
- Some silly person will try to benchmark your zfs web-API and overload
your server with zfs processes.
- Example: let me run [ ab -c 10 -n 10000 http://yourserver/zfs-api/list ]
-- Let me run 10 concurrent connection with a total of 10k requests to your
api (it's a simple one liner -- people will be tempted to benchmark like
this).

Example:
https://github.com/tlhakhan/ideal-potato/blob/master/zdux/routers/zfs/service.js#L9
- This is a JS example, but you can easily script it or another language
(golang) for cache separation and another program for the API.

Also, zfs does have a -c property to get cached values -- these values are
stored in an internal zfs process cache.  The -c doesn't help if you have
1000(0)s of filesystems, a single list can still take minutes.  Sending the
list is also several megabytes.

zfs list -Hrpc -o space

zfs get -Hrpc space all

- H= no headers
- r = recursive
- p = machine parseable
- c = hit cached entries

Fixes:  if ok, it may be good to stop the API, kill slowly the zfs list -t
all.

@ Eugene:
- I have seen single zfs list -Ho space -rt all take about 4-5 minutes, on
an 8000+ zfs_dataset zpool.

---
Notes:  my knowledge is from the illumos-zfs man pages but should apply
here.


On Sun, Aug 13, 2017 at 6:09 AM, Eugene M. Zheganin <emz@norma.perm.ru>
wrote:

> Hi,
>
> On 12.08.2017 20:50, Paul Kraus wrote:
>
>> On Aug 11, 2017, at 2:28 AM, Eugene M. Zheganin <emz@norma.perm.ru>
>>> wrote:
>>>
>>> Why does the zfs listing eat so much of the CPU ?
>>> 47114 root           1  20    0 40432K  3840K db->db  4   0:05 26.84% zfs
>>> 47099 root           1  20    0 40432K  3840K zio->i 17   0:05 26.83% zfs
>>> 47106 root           1  20    0 40432K  3840K db->db 21   0:05 26.81% zfs
>>> 47150 root           1  20    0 40432K  3428K db->db 13   0:03 26.31% zfs
>>> 47141 root           1  20    0 40432K  3428K zio->i 28   0:03 26.31% zfs
>>> 47135 root           1  20    0 40432K  3312K g_wait  9   0:03 25.51% zfs
>>> This is from winter 2017 11-STABLE (r310734), one of the 'zfs'es is
>>> cloning, and all the others are 'zfs list -t all'. I have like 25 gigs of
>>> free RAM, do I have any chance of speeding this up using may be some
>>> caching or some sysctl tuning ? We are using a simple ZFS web API that may
>>> issue concurrent or sequential listing requests, so as you can see they
>>> sometimes do stack.
>>>
>> How many snapshots do you have ? I have only seen this behavior with LOTS
>> (not hundreds, but thousands) of snapshots.
>>
> [root@san1:~]# zfs list -t snapshot | wc -l
>       88
>
>> What does your `iostat -x 1` look like ? I expect that you are probably
>> saturating your drives with random I/O.
>>
>
> Well, it's really long, and the disks are busy with random i/o indeed, but
> byst only for 20-30%. As about iostat - it's really long, because I have
> hundreds (not thousands) of zvols, and they do show up in iostat -x. But
> nothing unusual besides that.
>
>
> Thanks.
>
> Eugene.
>
>
> _______________________________________________
> freebsd-fs@freebsd.org mailing list
> https://lists.freebsd.org/mailman/listinfo/freebsd-fs
> To unsubscribe, send any mail to "freebsd-fs-unsubscribe@freebsd.org"
>



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?CALcn87zig7b37Q_JByVMrWyPPF=qr%2BWO%2BcB4aP4-wFp-_B4FOw>