Re: sysctl is too slow

From: Alan Somers <asomers_at_freebsd.org>
Date: Thu, 08 Sep 2022 23:30:31 UTC
I finally got around to this.  Here's the review:
https://reviews.freebsd.org/D36500

On Mon, Aug 23, 2021 at 6:54 AM Alan Somers <asomers@freebsd.org> wrote:
>
> Ideally, but it's not very high priority, since it's merely a performance issue in a monitoring tool.
>
> On Mon, Aug 23, 2021 at 6:05 AM Mateusz Guzik <mjguzik@gmail.com> wrote:
>>
>> So is this something you plan on fixing?
>>
>> On 8/17/21, Alan Somers <asomers@freebsd.org> wrote:
>> > Actually, I did get a flamegraph, and only 0.77% of samples were in ZFS.
>> >
>> > On Mon, Aug 16, 2021 at 7:19 PM Mateusz Guzik <mjguzik@gmail.com> wrote:
>> >
>> >> On 8/16/21, Alan Somers <asomers@freebsd.org> wrote:
>> >> > Yes, I see what you're talking about now.  There are a bunch of linked
>> >> > lists in sysctl_find_oid etc.  Good point.
>> >> > -Alan
>> >> >
>> >>
>> >> You still want to get a flamegraph, chances are most of the problem is in
>> >> zfs.
>> >>
>> >> > On Mon, Aug 16, 2021 at 1:30 PM Mateusz Guzik <mjguzik@gmail.com>
>> >> > wrote:
>> >> >
>> >> >> Last time I checked lookup of a sysctl was very bad with linear scans
>> >> all
>> >> >> over.
>> >> >>
>> >> >> Short of complete revamp of the entire thing I would start with
>> >> >> replacing the scans with a RB tree at each level. As is if you indeed
>> >> >> have 5000 datasets, you are doing increasingly longer walks.
>> >> >>
>> >> >> On 8/16/21, Alan Somers <asomers@freebsd.org> wrote:
>> >> >> > ztop feels very sluggish on a server with 5000 ZFS datasets.  Dtrace
>> >> >> shows
>> >> >> > that almost all of its time is spent in sys_sysctl.  ktrace shows
>> >> >> > that
>> >> >> both
>> >> >> > ztop and sysctl(8) call sys_sysctl a total of five times for each
>> >> >> > sysctl
>> >> >> > they care about:
>> >> >> >
>> >> >> > 1) To get the next oid
>> >> >> > 2) To get the sysctl's name
>> >> >> > 3) To get the oidfmt
>> >> >> > 4) To get the size of the value
>> >> >> > 5) To get the value itself.
>> >> >> >
>> >> >> > Each of these steps takes about equal time, and together all five
>> >> >> > take
>> >> >> > about 100us.  If the time per call is mostly syscall overhead, then
>> >> the
>> >> >> > process could be sped up by 80% by combining all of these things
>> >> >> > into
>> >> a
>> >> >> > single syscall: return the next oid, its name, its format, the size
>> >> >> > of
>> >> >> its
>> >> >> > value, and optimistically the value itself, assuming the user passed
>> >> >> > a
>> >> >> > sufficiently large buffer.
>> >> >> >
>> >> >> > Am I missing something?  Is there any other reason why sysctl is so
>> >> >> > slow?
>> >> >> > Or should I forget about it, and try to export ZFS's dataset stats
>> >> >> through
>> >> >> > devstat instead?
>> >> >> > -Alan
>> >> >> >
>> >> >>
>> >> >>
>> >> >> --
>> >> >> Mateusz Guzik <mjguzik gmail.com>
>> >> >>
>> >> >
>> >>
>> >>
>> >> --
>> >> Mateusz Guzik <mjguzik gmail.com>
>> >>
>> >
>>
>>
>> --
>> Mateusz Guzik <mjguzik gmail.com>