Poor netatalk performance through a bonded interface

Hi,

I am using pretty much the same configuration like described below, and everything works fine. My company provides a clustered storage solution, and we use samba and netatalk for CIFS and AFP support, and get really good performance, 50-60 MB/s in AFP and a bit slower in CIFS. We use 2 GB NICs for the client network, which are bonded in balance-alb mode, and atalkd is running flawlessly, broadcasting itself via DDP. However, I believe that all S X clients connect to the netatalk server using TCP (although we also support DDP), it must be default for them, and all S 9 connect to the server using DDP and still all works well.

There is one issue with afpd and atalkd, I don't know if it is related - afpd returns an IP address through DDP, but it is always the first IP address on the machine, i.e. usually on eth0. In our systems the client network is not necessarily on eth0, so we had to specify a different address specifically (using -hostname option) in afpd.conf.

Here are my afpd.conf, netatalk.conf and atalkd.conf:

cat /

# File: /

typhoon -closevol -cnidserver 10.57.0.2:33700 -hostname afp.typhoon -nosavepassword -nosetpassword -noslp -transall -uamlist uams_clrtxt.so,uams_dhx.so

cat /

# File: /

# Last changed by rnd on Thu Aug 11 08:13:53 2005

A2BT_RUN=no

AFPD_GUEST=nobody

AFPD_MAX_CLIENTS=200

AFPD_RUN=yes

ATALKD_RUN=yes

ATALK_BGRUND=no

ATALK_MAC_CHARSET='MAC_RMAN'

ATALK_NAME=`echo ${HSTNAME}|cut -d. -f1`

ATALK_UNIX_CHARSET='LCALE'

CNID_METAD_RUN=yes

PAPD_RUN=yes

TIMELRD_RUN=no

export ATALK_MAC_CHARSET

export ATALK_UNIX_CHARSET

ifconfig bond1

bond1 Link encap:Ethernet HWaddr 00:07:E9:17:D7:E6

inet addr:192.168.31.10 Bcast:192.168.31.255 Mask:255.255.255.0

EtherTalk Phase 2 addr:65280/146

UP BRADCAST RUNNING ALLMULTI MASTER MULTICAST MTU:1500 Metric:1

RX packets:3353444 errors:0 dropped:0 overruns:0 frame:0

TX packets:9484388 errors:0 dropped:0 overruns:0 carrier:0

collisions:0

RX bytes:1697676962 (1619.0 Mb) TX bytes:828293875 (789.9 Mb)

cat /

bond1 -phase 2 -net 0-65534 -addr 65280.146

Shlomi

From: netatalk-admins-admin (AT) lists (DOT) sourceforge.net on behalf of Rory Campbell-Lange

Sent: Wed 8/17/2005 11:20 PM

To: Mark Baker

Cc: netatalk-admins (AT) lists (DOT) sourceforge.net

Subject: Re: [Netatalk-admins] Poor netatalk performance through a bonded interface

Hi Mark

Check out my email from a while ago (quoted below). I couldn't find a

solution I'm sorry to say.

Regards,

Rory

02/08/05, Mark Baker (mark.baker (AT) silicon (DOT) co.uk) wrote:

I am running RH8 with a 2.4.28 kernel, using the bonding drive to bond 2 x

GB NICS together.

Firstly, DDP does not work through the bonded interface, so I have turned

this off using ?noddp in my afpd.conf file, but also had to turn off ATALKD

in netatalk.conf.

Secondly, connecting from an SX 10.4.2 machine via afp to the IP address,

connects me, and browsing the file system is quick, however, transferring

files is painfully slow, were talking .5mb a second here.

From: Rory Campbell-Lange <rory (AT) campbell-lange (DOT) net>

To: Mike Weller <test3 (AT) zyvex (DOT) com>

Cc:

Bcc:

Subject: Re: [Netatalk-devel] atalk doesnt like bond0 interface

Reply-To:

In-Reply-To: <20020211100458.GA1867 (AT) campbell-lange (DOT) net>

11/02/02, Rory Campbell-Lange (rory (AT) campbell-lange (DOT) net) wrote:

Hi Mike.

I have had similar problems supporting bonding (I'm using ifenslave to

enslave the interfaces to eth0). It appears that the kernel support for

appletalk doesn't support a bonded interface. I haven't come across any

solution; as a result our main server runs on tcp only.

I don't have any links on this, but there are plenty of historical

complaints to be found on Google!

07/02/02, Mike wrote:

Now, when I turn channel-bonding on with:

/sbin/insmod bonding

/sbin/ifconfig bond0 inet 10.0.0.18 netmask 255.255.0.0

/sbin/ifconfig eth0 inet 10.0.0.18 netmask 255.255.0.0

/sbin/ifconfig eth1 inet 10.0.0.18 netmask 255.255.0.0

/sbin/ifenslave bond0 eth0

/sbin/ifenslave bond0 eth1

(3rd and 4th line probably arent necessary, but it doesnt hurt)

I set up bond0 aliases to stop the interface hopping to eth0. You will see

that the three network interfaces share the network load the addition of

the normal interface names was to try and get around appletalk not

working with bond0. Butit hasn't worked.

I have:

bond0 Link encap:Ethernet HWaddr 00:D0:B7:A9:89:5F

inet addr:10.0.0.50 Bcast:10.255.255.255 Mask:255.0.0.0

UP BRADCAST RUNNING MASTER MULTICAST MTU:1500 Metric:1

RX packets:0 errors:0 dropped:0 overruns:0 frame:0

TX packets:809082955 errors:0 dropped:403459 overruns:0 carrier:0

collisions:0

bond0:1 Link encap:Ethernet HWaddr 00:D0:B7:A9:89:5F

inet addr:10.0.0.51 Bcast:10.255.255.255 Mask:255.0.0.0

UP BRADCAST RUNNING MASTER MULTICAST MTU:1500 Metric:1

bond0:2 Link encap:Ethernet HWaddr 00:D0:B7:A9:89:5F

inet addr:10.0.0.52 Bcast:10.255.255.255 Mask:255.0.0.0

UP BRADCAST RUNNING MASTER MULTICAST MTU:1500 Metric:1

eth0 Link encap:Ethernet HWaddr 00:D0:B7:A9:89:5F

inet addr:10.0.0.50 Bcast:10.255.255.255 Mask:255.0.0.0

EtherTalk Phase 2 addr:65280/214 (Helios Ethershare interface)

UP BRADCAST RUNNING ALLMULTI SLAVE MULTICAST MTU:1500 Metric:1

RX packets:206672186 errors:2495 dropped:0 overruns:270959 frame:0

TX packets:269695274 errors:0 dropped:0 overruns:137 carrier:1019500

collisions:1137539

eth1 Link encap:Ethernet HWaddr 00:D0:B7:A9:89:5F

inet addr:10.0.0.50 Bcast:10.255.255.255 Mask:255.0.0.0

UP BRADCAST RUNNING SLAVE MULTICAST MTU:1500 Metric:1

RX packets:245774589 errors:248 dropped:0 overruns:6017 frame:1

TX packets:269694319 errors:0 dropped:0 overruns:153 carrier:1591086

collisions:1729205

eth2 Link encap:Ethernet HWaddr 00:D0:B7:A9:89:5F

inet addr:10.0.0.50 Bcast:10.255.255.255 Mask:255.0.0.0

UP BRADCAST RUNNING SLAVE MULTICAST MTU:1500 Metric:1

RX packets:273469154 errors:110 dropped:0 overruns:6260 frame:2

TX packets:269694711 errors:0 dropped:0 overruns:153 carrier:3999234

collisions:4168778

Rory

http://network.itags.org/q_networking-tech_184696.html