Date: Sun, 16 Nov 2008 03:18:46 -0800 (PST) From: Won De Erick <won.derick@yahoo.com> To: Jeremy Chadwick <koitsu@FreeBSD.org> Cc: freebsd-hackers@freebsd.org, rwatson@freebsd.org Subject: Re: NET.ISR and CPU utilization performance w/ HP DL 585 using FreeBSD 7.1 Beta2 Message-ID: <193147.59821.qm@web45803.mail.sp1.yahoo.com> References: <557765.55617.qm@web45804.mail.sp1.yahoo.com> <20081115141631.GA75733@icarus.home.lan>
next in thread | previous in thread | raw e-mail | index | archive | help
> ----- Original Message ----
> From: Jeremy Chadwick <koitsu@FreeBSD.org>
> To: Won De Erick <won.derick@yahoo.com>
> Cc: rwatson@freebsd.org; freebsd-hackers@freebsd.org
> Sent: Saturday, November 15, 2008 10:16:31 PM
> Subject: Re: NET.ISR and CPU utilization performance w/ HP DL 585 using FreeBSD 7.1 Beta2
>
> On Sat, Nov 15, 2008 at 04:59:16AM -0800, Won De Erick wrote:
> > Hello,
> >
> > I tested HP DL 585 (16 CPUs, w/ built-in Broadcom NICs) running FreeBSD 7.1 Beta2 under heavy network traffic (TCP).
> >
> > SCENARIO A : Bombarded w/ TCP traffic:
> >
> > When net.isr.direct=1,
> >
> > PID USERNAME THR PRI NICE SIZE RES STATE C TIME WCPU COMMAND
> > 52 root 1 -68 - 0K 16K CPU11 b 38:43 95.36% irq32: bce1
> > 51 root 1 -68 - 0K 16K CPU10 a 25:50 85.16% irq31: bce0
> > 16 root 1 171 ki31 0K 16K RUN a 65:39 15.97% idle: cpu10
> > 28 root 1 -32 - 0K 16K WAIT 8 12:28 5.18% swi4: clock sio
> > 15 root 1 171 ki31 0K 16K RUN b 52:46 3.76% idle: cpu11
> > 45 root 1 -64 - 0K 16K WAIT 7 7:29 1.17% irq17: uhci0
> > 47 root 1 -64 - 0K 16K WAIT 6 1:11 0.10% irq16: ciss0
> > 27 root 1 -44 - 0K 16K WAIT 0 28:52 0.00% swi1: net
> >
> > When net.isr.direct=0,
> >
> > 16 root 1 171 ki31 0K 16K CPU10 a 106:46 92.58% idle: cpu10
> > 19 root 1 171 ki31 0K 16K CPU7 7 133:37 89.16% idle: cpu7
> > 27 root 1 -44 - 0K 16K WAIT 0 52:20 76.37% swi1: net
> > 25 root 1 171 ki31 0K 16K RUN 1 132:30 70.26% idle: cpu1
> > 26 root 1 171 ki31 0K 16K CPU0 0 111:58 64.36% idle: cpu0
> > 15 root 1 171 ki31 0K 16K CPU11 b 81:09 57.76% idle: cpu11
> > 52 root 1 -68 - 0K 16K WAIT b 64:00 42.97% irq32: bce1
> > 51 root 1 -68 - 0K 16K WAIT a 38:22 12.26% irq31: bce0
> > 45 root 1 -64 - 0K 16K WAIT 7 11:31 12.06% irq17: uhci0
> > 47 root 1 -64 - 0K 16K WAIT 6 1:54 3.66% irq16: ciss0
> > 28 root 1 -32 - 0K 16K WAIT 8 16:01 0.00% swi4: clock sio
> >
> > Overall CPU utilization has significantly dropped, but I noticed that swi1 has taken CPU0 with high utilization when the net.isr.direct=0.
> > What does this mean?
> >
> > SCENARIO B : Bombarded w/ more TCP traffic:
> >
> > Worst thing, the box has become unresponsive (can't be PINGed, inaccessible through SSH) after more traffic was added retaining net.isr.direct=0.
> > This is due maybe to the 100% utilization on CPU0 for sw1:net (see below result, first line). bce's and swi's seem to race each other based on the result when net.isr.direct=1, swi1 .
> > The rest of the CPUs are sitting pretty (100% Idle). Can you shed some lights on this?
> >
> > When net.isr.direct=0:
> > 27 root 1 -44 - 0K 16K CPU0 0 5:45 100.00% swi1: net
> > 11 root 1 171 ki31 0K 16K CPU15 0 0:00 100.00% idle: cpu15
> > 13 root 1 171 ki31 0K 16K CPU13 0 0:00 100.00% idle: cpu13
> > 17 root 1 171 ki31 0K 16K CPU9 0 0:00 100.00% idle: cpu9
> > 18 root 1 171 ki31 0K 16K CPU8 0 0:00 100.00% idle: cpu8
> > 21 root 1 171 ki31 0K 16K CPU5 5 146:17 99.17% idle: cpu5
> > 22 root 1 171 ki31 0K 16K CPU4 4 146:17 99.07% idle: cpu4
> > 14 root 1 171 ki31 0K 16K CPU12 0 0:00 99.07% idle: cpu12
> > 16 root 1 171 ki31 0K 16K CPU10 a 109:33 98.88% idle: cpu10
> > 15 root 1 171 ki31 0K 16K CPU11 b 86:36 93.55% idle: cpu11
> > 52 root 1 -68 - 0K 16K WAIT b 59:42 13.87% irq32: bce1
> >
> > When net.isr.direct=1,
> > 52 root 1 -68 - 0K 16K CPU11 b 55:04 97.66% irq32: bce1
> > 51 root 1 -68 - 0K 16K CPU10 a 33:52 73.88% irq31: bce0
> > 16 root 1 171 ki31 0K 16K RUN a 102:42 26.86% idle: cpu10
> > 15 root 1 171 ki31 0K 16K RUN b 81:20 3.17% idle: cpu11
> > 28 root 1 -32 - 0K 16K WAIT e 13:40 0.00% swi4: clock sio
> >
> > With regards to bandwidth in all scenarios above, the result is extremely low (expected is several hundred Mb/s). Why?
The below result should be under scenario B above only.
> >
> > - iface Rx Tx Total
> > ==============================================================================
> > bce0: 4.69 Mb/s 10.49 Mb/s 15.18 Mb/s
> > bce1: 20.66 Mb/s 4.68 Mb/s 25.34 Mb/s
> > lo0: 0.00 b/s 0.00 b/s 0.00 b/s
> > ------------------------------------------------------------------------------
> > total: 25.35 Mb/s 15.17 Mb/s 40.52 Mb/s
> >
> >
> > Thanks,
> >
> > Won
>
> And does this behaviour change if you use some other brand of NIC?
With Intel Pro NIC ( 82571):
When net.isr.direct=1,
49 root 1 -68 - 0K 16K CPU12 c 6:50 100.00% em0 taskq
15 root 1 171 ki31 0K 16K CPU11 b 5:47 100.00% idle: cpu11
50 root 1 -68 - 0K 16K CPU13 d 6:15 86.96% em1 taskq
25 root 1 171 ki31 0K 16K CPU1 1 9:27 79.79% idle: cpu1
28 root 1 -32 - 0K 16K WAIT 1 1:33 22.75% swi4: clock sio
13 root 1 171 ki31 0K 16K RUN d 4:14 12.26% idle: cpu13
14 root 1 171 ki31 0K 16K RUN c 3:37 0.00% idle: cpu12
em0 and em1 have high CPU utilization, and with netstat, there were packet errors.
# netstat -I em0 -w 1 -d
input (em0) output
packets errs bytes packets errs bytes colls drops
15258 3066 22748316 18468 0 4886567 0 0
15461 3096 22783724 18379 0 5350130 0 0
When net.isr.direct=0,
12 root 1 171 ki31 0K 16K CPU14 e 22:28 100.00% idle: cpu14
20 root 1 171 ki31 0K 16K CPU6 6 24:32 97.85% idle: cpu6
25 root 1 171 ki31 0K 16K RUN 1 21:51 96.97% idle: cpu1
27 root 1 -44 - 0K 16K CPU2 2 5:12 91.55% swi1: net
13 root 1 171 ki31 0K 16K CPU13 d 11:04 86.96% idle: cpu13
14 root 1 171 ki31 0K 16K CPU12 c 10:51 81.59% idle: cpu12
49 root 1 -68 - 0K 16K CPU12 c 13:48 22.17% em0 taskq
24 root 1 171 ki31 0K 16K RUN 2 19:16 12.16% idle: cpu2
50 root 1 -68 - 0K 16K - d 13:34 11.87% em1 taskq
28 root 1 -32 - 0K 16K WAIT 3 3:48 0.00% swi4: clock sio
sw1:net is taking high CPU utilization this time, but without packet errors:
# netstat -I em0 -w 1 -d
input (em0) output
packets errs bytes packets errs bytes colls drops
4275 0 5528012 24878 0 24162198 0 0
4317 0 5585954 24880 0 24066583 0 0
Is this related to the context switching in FreeBSD 7.x? I noticed that there were no significant difference in enabling and disabling net.isr.direct in FreeBSD 6.2.
Also, is there any significance of enabling device polling?
>
> --
> | Jeremy Chadwick jdc at parodius.com |
> | Parodius Networking http://www.parodius.com/ |
> | UNIX Systems Administrator Mountain View, CA, USA |
> | Making life hard for others since 1977. PGP: 4BD6C0CB |
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?193147.59821.qm>