Skip site navigation (1)Skip section navigation (2)
Date:      Mon, 15 Oct 2018 17:05:54 -0400
From:      James Shank <james.shank@gmail.com>
To:        freebsd-hackers@freebsd.org
Subject:   High I/O RAID on UFS2
Message-ID:  <1BA62E58-93D7-4763-B27A-E21B7DA362B2@gmail.com>

next in thread | raw e-mail | index | archive | help
Hi FreeBSD Hackers,

I am running FreeBSD 11.1-RELEASE-p15 with a UFS2 filesystem block size 64K,=
 with a RAID 6 of 16 SSD drives on a MRSAS compatible RAID card stripe size o=
f 256K (and yes, I know about the misalignment) and 384GB memory on 48 real c=
ores with hyper threading enabled.  I invite critiques of this configuration=
, but am more keen on getting help making it run efficiently.

I am running 11.1 because 11.2-RELEASE repeatably panics after failing somew=
here after allocbuf() tries to allocate a page (sorry, I don=E2=80=99t still=
 have the output from this state).

I am doing very read heavy operations, via multithreaded readers.  My writes=
 tend to be small, but come in every 15 seconds or so.  I=E2=80=99d estimate=
 read to write ratio of about 100:1 or more.

What I see happening is pagedaemon coming alive and bufferspacedaemon doing t=
he same to clear the corresponding resources.  bufferspacedaemon runs nearly=
 all the time at 100%.  I haven=E2=80=99t yet had much success with tuning t=
hings to keep enough resources free to allow my read rates to be sustained.

My read rates can hit and exceed 6GBps when everything is humming along, but=
 only for short periods of time, then they drop off to much lower rates afte=
r the kernel starts shuffling resources.  For long periods of time, my best s=
ustainable rate seems to be 3.5GBps, and I can=E2=80=99t seem to get the ful=
l read rate for sustained periods.

I have tried several things, but I believe now that I may be missing somethi=
ng that the more experienced gurus will find obvious.

Any pointers on where I should look next?

Thanks!

James



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?1BA62E58-93D7-4763-B27A-E21B7DA362B2>