Date: Wed, 20 Aug 2014 10:56:56 -0500 From: "Polyack, Steve" <Steve.Polyack@intermedix.com> To: Alan Cox <alc@rice.edu>, "freebsd-stable@freebsd.org" <freebsd-stable@freebsd.org> Subject: RE: vmdaemon CPU usage and poor performance in 10.0-RELEASE Message-ID: <4D557EC7CC2A544AA7C1A3B9CBA2B3672609CF335D@exchange03.epbs.com> In-Reply-To: <53F4C4C2.1030109@rice.edu> References: <4D557EC7CC2A544AA7C1A3B9CBA2B36726098846B4@exchange03.epbs.com> <20140813152522.GI9400@home.opsec.eu> <4D557EC7CC2A544AA7C1A3B9CBA2B36726098847AF@exchange03.epbs.com> <CAJUyCcNoTJ3xqkC_Prz3N%2BApEqYy3Mi2gA%2BuDo33dczaTMONrA@mail.gmail.com> <4D557EC7CC2A544AA7C1A3B9CBA2B3672609BBA3C4@exchange03.epbs.com> <53F24E5B.1010809@rice.edu> <4D557EC7CC2A544AA7C1A3B9CBA2B3672609BBA64F@exchange03.epbs.com> <53F2790C.20703@rice.edu> <4D557EC7CC2A544AA7C1A3B9CBA2B3672609CF28E5@exchange03.epbs.com> <4D557EC7CC2A544AA7C1A3B9CBA2B3672609CF2F8F@exchange03.epbs.com> <4D557EC7CC2A544AA7C1A3B9CBA2B3672609CF31F0@exchange03.epbs.com> <53F4C4C2.1030109@rice.edu>
next in thread | previous in thread | raw e-mail | index | archive | help
> -----Original Message-----
> From: Alan Cox [mailto:alc@rice.edu]
> Sent: Wednesday, August 20, 2014 11:55 AM
> To: Polyack, Steve; freebsd-stable@freebsd.org
> Subject: Re: vmdaemon CPU usage and poor performance in 10.0-RELEASE
>=20
> On 08/20/2014 09:55, Polyack, Steve wrote:
> >> -----Original Message-----
> >> From: Polyack, Steve
> >> Sent: Wednesday, August 20, 2014 9:14 AM
> >> To: Polyack, Steve; Alan Cox; freebsd-stable@freebsd.org
> >> Subject: RE: vmdaemon CPU usage and poor performance in 10.0-RELEASE
> >>
> >>
> >>> -----Original Message-----
> >>> From: owner-freebsd-stable@freebsd.org [mailto:owner-freebsd-
> >>> stable@freebsd.org] On Behalf Of Polyack, Steve
> >>> Sent: Tuesday, August 19, 2014 12:37 PM
> >>> To: Alan Cox; freebsd-stable@freebsd.org
> >>> Subject: RE: vmdaemon CPU usage and poor performance in 10.0-
> RELEASE
> >>>
> >>>> -----Original Message-----
> >>>> From: owner-freebsd-stable@freebsd.org [mailto:owner-freebsd-
> >>>> stable@freebsd.org] On Behalf Of Alan Cox
> >>>> Sent: Monday, August 18, 2014 6:07 PM
> >>>> To: freebsd-stable@freebsd.org
> >>>> Subject: Re: vmdaemon CPU usage and poor performance in 10.0-
> >> RELEASE
> >>>> On 08/18/2014 16:29, Polyack, Steve wrote:
> >>>>>> -----Original Message-----
> >>>>>> From: owner-freebsd-stable@freebsd.org [mailto:owner-freebsd-
> >>>>>> stable@freebsd.org] On Behalf Of Alan Cox
> >>>>>> Sent: Monday, August 18, 2014 3:05 PM
> >>>>>> To: freebsd-stable@freebsd.org
> >>>>>> Subject: Re: vmdaemon CPU usage and poor performance in 10.0-
> >>> RELEASE
> >>>>>> On 08/18/2014 13:42, Polyack, Steve wrote:
> >>>>>>> Excuse my poorly formatted reply at the moment, but this seems to
> >>>> have
> >>>>>> fixed our problems. I'm going to update the bug report with a not=
e.
> >>>>>>> Thanks Alan!
> >>>>>> You're welcome. And, thanks for letting me know of the outcome.
> >>>>>>
> >>>>> Actually, I may have spoken too soon, as it looks like we're seeing
> >>>> vmdaemon tying up the system again:
> >>>>> root 6 100.0 0.0 0 16 - DL Wed04P=
M 4:37.95
> >>> [vmdaemon]
> >>>>> Is there anything I can check to help narrow down what may be the
> >>>> problem? KTrace/truss on the "process" doesn't give any information=
, I
> >>>> suppose because it's actually a kernel thread.
> >>>>
> >>>> Can you provide the full output of top? Is there anything unusual
> about
> >>>> the hardware or software configuration?
> >>> This may have just been a fluke (maybe NFS caching the old
> vm_pageout.c
> >>> during the first source build). We've rebuilt and are monitoring it =
now.
> >>>
> >>> The hardware consists of a few Dell PowerEdge R720xd servers with
> 256GB
> >>> of RAM and array of SSDs (no ZFS). 64GB is dedicated to postgres
> >>> shared_buffers right now. FreeBSD 10, PostgreSQL 9.3, Slony-I v2.2.2,
> and
> >>> redis-2.8.11 are all in use here. I can't say that anything is unusu=
al about
> >> the
> >>> configuration.
> >>>
> >> We are still seeing the issue. It seems to manifest once the "Free"
> memory
> >> gets under 10GB (of 256GB on the system), even though ~200GB of this i=
s
> >> classified as Inactive. For us, this was about 7 hours of database ac=
tivity
> >> (initial replication w/ slony). Right now vmdaemon is consuming 100%
> CPU
> >> and shows 671:34 CPU time when it showed 0:00 up until the problem
> >> manifested. The full top output (that fits on my screen) is below:
> >>
> >> last pid: 62309; load averages: 4.05, 4.24, 4.10
> >> up 0+22:34:31 09:08:43
> >> 159 processes: 8 running, 145 sleeping, 1 waiting, 5 lock
> >> CPU: 14.5% user, 0.0% nice, 4.9% system, 0.0% interrupt, 80.5% idle
> >> Mem: 26G Active, 216G Inact, 4122M Wired, 1178M Cache, 1632M Buf,
> 2136M
> >> Free
> >> Swap: 32G Total, 32G Free
> >>
> >> PID USERNAME THR PRI NICE SIZE RES STATE C TIME WC=
PU
> >> COMMAND
> >> 11 root 32 155 ki31 0K 512K CPU31 31 669.6H 2934.=
23% idle
> >> 6 root 1 -16 - 0K 16K CPU19 19 678:57 100.0=
0% vmdaemon
> >> 1963 pgsql 1 45 0 67538M 208M CPU0 0 121:46 17.3=
8% postgres
> >> 2037 pgsql 1 77 0 67536M 2200K *vm ob 14 6:24 15.9=
7% postgres
> >> 1864 pgsql 1 31 0 67536M 1290M semwai 4 174:41 15.1=
9%
> postgres
> >> 1996 pgsql 1 38 0 67538M 202M semwai 16 120:27 15.0=
9%
> postgres
> >> 1959 pgsql 1 39 0 67538M 204M CPU27 27 117:30 15.0=
9% postgres
> >> 1849 pgsql 1 32 0 67536M 1272M semwai 23 126:22 13.9=
6%
> postgres
> >> 1997 pgsql 1 31 0 67538M 206M CPU30 30 122:26 11.7=
7% postgres
> >> 2002 pgsql 1 34 0 67538M 182M sbwait 11 55:20 11.2=
8% postgres
> >> 1961 pgsql 1 32 0 67538M 206M CPU12 12 121:47 10.9=
9% postgres
> >> 1964 pgsql 1 30 0 67538M 206M semwai 28 122:08 9.8=
6% postgres
> >> 1962 pgsql 1 29 0 67538M 1286M sbwait 2 45:49 7.1=
8% postgres
> >> 1752 root 1 22 0 78356K 8688K CPU2 2 175:46 6.8=
8% snmpd
> >> 1965 pgsql 1 25 0 67538M 207M semwai 9 120:55 6.5=
9% postgres
> >> 1960 pgsql 1 23 0 67538M 177M semwai 6 52:42 4.8=
8% postgres
> >> 1863 pgsql 1 25 0 67542M 388M semwai 25 9:12 2.2=
0% postgres
> >> 1859 pgsql 1 22 0 67538M 1453M *vm ob 20 6:13 2.1=
0% postgres
> >> 1860 pgsql 1 22 0 67538M 1454M sbwait 8 6:08 1.9=
5% postgres
> >> 1848 pgsql 1 21 0 67586M 66676M *vm ob 30 517:07 1.6=
6%
> postgres
> >> 1856 pgsql 1 22 0 67538M 290M *vm ob 15 5:39 1.6=
6% postgres
> >> 1846 pgsql 1 21 0 67538M 163M sbwait 15 5:46 1.4=
6% postgres
> >> 1853 pgsql 1 21 0 67538M 110M sbwait 30 8:54 1.1=
7% postgres
> >> 1989 pgsql 1 23 0 67536M 5180K sbwait 18 1:41 0.9=
8% postgres
> >> 5 root 1 -16 - 0K 16K psleep 6 9:33 0.7=
8% pagedaemon
> >> 1854 pgsql 1 20 0 67538M 338M sbwait 22 5:38 0.7=
8% postgres
> >> 1861 pgsql 1 20 0 67538M 286M sbwait 15 6:13 0.6=
8% postgres
> >> 1857 pgsql 1 20 0 67538M 1454M semwai 10 6:19 0.4=
9% postgres
> >> 1999 pgsql 1 36 0 67538M 156M *vm ob 28 120:56 0.3=
9% postgres
> >> 1851 pgsql 1 20 0 67538M 136M sbwait 22 5:48 0.3=
9% postgres
> >> 1975 pgsql 1 20 0 67536M 5688K sbwait 25 1:40 0.2=
9% postgres
> >> 1858 pgsql 1 20 0 67538M 417M sbwait 3 5:55 0.2=
0% postgres
> >> 2031 pgsql 1 20 0 67536M 5664K sbwait 5 3:26 0.1=
0% postgres
> >> 1834 root 12 20 0 71892K 12848K select 20 34:05 0.0=
0% slon
> >> 12 root 78 -76 - 0K 1248K WAIT 0 25:47 0.0=
0% intr
> >> 2041 pgsql 1 20 0 67536M 5932K sbwait 14 12:50 0.0=
0% postgres
> >> 2039 pgsql 1 20 0 67536M 5960K sbwait 17 9:59 0.0=
0% postgres
> >> 2038 pgsql 1 20 0 67536M 5956K sbwait 6 8:21 0.0=
0% postgres
> >> 2040 pgsql 1 20 0 67536M 5996K sbwait 7 8:20 0.0=
0% postgres
> >> 2032 pgsql 1 20 0 67536M 5800K sbwait 22 7:03 0.0=
0% postgres
> >> 2036 pgsql 1 20 0 67536M 5748K sbwait 23 6:38 0.0=
0% postgres
> >> 1812 pgsql 1 20 0 67538M 59185M select 1 5:46 0.0=
0% postgres
> >> 2005 pgsql 1 20 0 67536M 5788K sbwait 23 5:14 0.0=
0% postgres
> >> 2035 pgsql 1 20 0 67536M 4892K sbwait 18 4:52 0.0=
0% <postgres>
> >> 1852 pgsql 1 21 0 67536M 1230M semwai 7 4:47 0.0=
0% postgres
> >> 13 root 3 -8 - 0K 48K - 28 4:46 0.0=
0% geom
> >>
> >>
> > Another thing I've noticed is that this sysctl vm.stats counter is incr=
easing
> fairly rapidly:
> > # sysctl vm.stats.vm.v_pdpages && sleep 1 && sysctl
> vm.stats.vm.v_pdpages
> > vm.stats.vm.v_pdpages: 3455264541
> > vm.stats.vm.v_pdpages: 3662158383
>=20
> I'm not sure what that tells us, because both the page daemon and the vm
> ("swap") daemon increment this counter.
>=20
> > Also, to demonstrate what kind of problems this seems to cause:
> > # time sleep 1
> >
> > real 0m18.288s
> > user 0m0.001s
> > sys 0m0.004s
>=20
> If you change the sysctl vm.swap_enabled to 0, how does your system
> behave?
>=20
Setting vm.swap_enabled to 0 made the problem clear up almost instantly. v=
mdaemon is back to 0.00% CPU usage and the system is responsive once again.
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?4D557EC7CC2A544AA7C1A3B9CBA2B3672609CF335D>