Date: Wed, 26 Feb 2003 22:05:51 +0900 (KST) From: CHOI Junho <cjh@kr.FreeBSD.org> To: net@freebsd.org Subject: Performance tuning hints of gigabit networking? Message-ID: <20030226.220551.10329540.cjh@kr.FreeBSD.org>
next in thread | raw e-mail | index | archive | help
Hi,
I am looking for a good resource for kernel tuning on very high
bandwidth HTTP servers(avg 500Mbit/sec, peak 950Mbit/sec). Today I
faced very unusual situation with 950Mbit/sec bandwidth!
> netstat -m
16962/93488/262144 mbufs in use (current/peak/max):
16962 mbufs allocated to data
16952/65536/65536 mbuf clusters in use (current/peak/max)
154444 Kbytes allocated to network (14% of mb_map in use)
512627 requests for memory denied
2614 requests for memory delayed
0 calls to protocol drain routines
I set kern.ipc.nmbclusters=65536, but it overflowed. This is P-IV Xeon
1.8G, 2GB RAM, and one Intel 1000baseSX(em driver) machine running
4.7-RELEASE-pX. This server is running only one service, HTTP. I use
thttpd, since apache doesn't work in such a high load. thttpd is highly
amazing, just give <1 load in any time.
Once I tried to increase kern.ipc.nmbclusters to 131072 or
higher(multiple of 65536 or 32768, tuning(7) only cites about 32768
case..), it fails to boot kernel when 262144, or kernel panic in
somewhat higher load when 131072, so I gave up other changes and fall
back to 65536.
What is a good way to calcurate this value safely? Here is another
hint, /etc/sysctl.conf:
net.link.ether.inet.log_arp_wrong_iface=0
kern.ipc.maxsockbuf=2048000
kern.ipc.somaxconn=4096
kern.ipc.maxsockets=60000
kern.maxfiles=65536
kern.maxfilesperproc=32768
net.inet.tcp.rfc1323=1
net.inet.tcp.delayed_ack=0
net.inet.tcp.sendspace=65535
net.inet.tcp.recvspace=65535
net.inet.udp.recvspace=65535
net.inet.udp.maxdgram=57344
net.inet.icmp.drop_redirect=1
net.inet.icmp.log_redirect=1
net.inet.ip.redirect=0
net.inet6.ip6.redirect=0
net.link.ether.inet.max_age=1200
net.inet.ip.sourceroute=0
net.inet.ip.accept_sourceroute=0
net.inet.icmp.bmcastecho=0
net.inet.icmp.maskrepl=0
net.inet.tcp.inflight_enable=1
kernel configuration is not specially tuned, except DEVICE_POLLING and
HZ=2000.
--
CHOI Junho <http://www.kr.FreeBSD.org/~cjh>; KFUG <cjh at kr.FreeBSD.org>
FreeBSD Project <cjh at FreeBSD.org> Web Data Bank <cjh at wdb.co.kr>
Key fingerprint = 1369 7374 A45F F41A F3C0 07E3 4A01 C020 E602 60F5
To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-net" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20030226.220551.10329540.cjh>