Skip site navigation (1)Skip section navigation (2)
Date:      Wed, 5 May 2004 16:34:28 -0500
From:      "adp" <dap99@i-55.com>
To:        "Charles Swiger" <cswiger@mac.com>
Cc:        performance@freebsd.org
Subject:   Re: Abnormal  network errors?
Message-ID:  <015e01c432e9$1d77b650$6501a8c0@yourqqh4336axf>
References:  <01a501c432ce$adf20e30$4b0a000a@yourqqh4336axf> <207CE1A8-9ECB-11D8-8DD7-003065ABFD92@mac.com>

next in thread | previous in thread | raw e-mail | index | archive | help
----- Original Message -----
From: "Charles Swiger" <cswiger@mac.com>

> On May 5, 2004, at 2:27 PM, adp wrote:
> > On this server I'm thinking I need two things:
> >
> > 1. More sockets available.
> > 2. Larger sockbufs for send and recv.
> >
> > Is this an accurate assessment?
>
> Given the application of this system, you might want to up the value of
> kern.ipc.nmbclusters by a factor of four or so (it's NBMCLUSTERS in the
> kernel config file).  However, it's not essential-- your "netstat -m"
> is OK, and your TCP send and receive windows are reasonably sized as-is
> by default.

Several problems. First, we are hosting a DNS server on this box. The DNS
resolves domains we are auth. for very fast, or anything in its cache very
fast, but anything else is SLOW or times-out. Also, our www server (another
box) is responding slowly in general (4-6 seconds).

> > What is "2432320 packets for unknown/unsupported protocol"? What
> > specifically does this mean? (In other words, what should I do to
> > resolve
> > this?)
>
> It means machines are sending non-IP traffic on your network, which is
> normal if you have Windows protocols (NetBEUI, SPX/IPX) or Macs
> (AppleTalk) around.  Or chatty network devices like some printers....

What is 802.1d? I am getting a lot of this:

16:21:16.788617 802.1d config 8000.00:04:27:d1:cb:d3.8019 root
8000.00:03:6c:51:a2:a7 pathcost 8 age 2 max 20 hello 2 fdelay 15

And this:

16:21:15.424508 CDP v2, ttl=180s DevID 'Six2' Addr (1): IPv4 10.2.254.62
PortID 'FastEthernet0/12' CAP 0x0a[|cdp]

I'm at a colo.

> See /usr/include/net/ethernet.h for an idea, or maybe "tcpdump not ip"
> might give some idea of what's going by.
>
> > What about "921363 calls to icmp_error"?
>
> ICMP messages like responding to a ping, or people sending traffic with
> RFC-1918 unroutable addresses (gives "dest unreachable")...

That's weird. I tried 'tcpdump icmp' and see a few errors right off the bat:

# tcpdump -n icmp
tcpdump: listening on rl0
16:27:46.633262 192.168.42.71 > 192.168.42.76: icmp: 192.168.42.71 udp port
1249 unreachable
16:27:53.639237 192.168.42.71 > 192.168.42.76: icmp: 192.168.42.71 udp port
1280 unreachable
16:28:02.579417 192.168.42.71 > 192.168.42.76: icmp: 192.168.42.71 udp port
1204 unreachable
16:28:07.716527 192.168.42.71 > 192.168.42.76: icmp: 192.168.42.71 udp port
1510 unreachable
16:28:08.589910 192.168.42.71 > 192.168.42.76: icmp: 192.168.42.71 udp port
1218 unreachable
16:28:15.668697 192.168.42.71 > 192.168.42.76: icmp: 192.168.42.71 udp port
1327 unreachable
16:28:33.581427 192.168.42.71 > 192.168.42.76: icmp: 192.168.42.71 udp port
1355 unreachable

Hmm. I am running a DNS server in a jail on the NFS server. We have been
getting very slow responses times from it. Seems related.

> > Under tcp I have "481930 embryonic connections dropped". I assume that
> > means
> > I don't have enough sockets available for when this server gets loaded.
> > Correct?
>
> More likely, these are someone doing a port scan and leaving half-open
> connections lying around to get cleaned up.

We are behind a managed NetScreen firewall. I can't see how anyone is
port-scanning us, unless they are just scanning the few ports we have open
to the world.

> It might be helpful if you gave us some idea as to what the performance
> problem you were seeing was?  Is NFS access slow, or some such?  Are
> you seeing errors or collisions in netstat -i or in whatever statistics
> your switch keeps per port?
>
> The following areas struck me as being relevant:
>
> > # ifconfig rl0
>
> First, consider upgrading to a fxp or dc-based NIC.

Noted.

> > udp:
> >         272987897 datagrams received
> > [ ... ]
> >         19976574 dropped due to full socket buffers
>
> This is high enough to represent a concern, agreed.

How do I fix this then? I assume I don't have enough sockets available. Is
there a way to see where I am peaking?

I'm thinking adding more memory will increase the system-set defaults (also
read 'man tuning').

> > ip:
> >         578001924 total packets received
> [ ... ]
> >         4899083 fragments received
> >         4 fragments dropped (dup or out of space)
> >         750 fragments dropped after timeout
> >         842689 packets reassembled ok
> [ ... ]
> >         609745425 packets sent from this host
> >         1914687 output datagrams fragmented
> >         10496350 fragments created
>
> Second, you're fragmenting a relatively large number of packets going
> by, you ought to see what's going on with your MTU and pMTU discovery.
> I suppose if you're using large UDP datagrams with NFS, that might be
> it.
>
> [ The machines I've got around with comparible traffic volume might
> have 400 frags received, and 10 transmitted, or some such. ]

Could this be related to my switch or anything else?





Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?015e01c432e9$1d77b650$6501a8c0>