Dear list people,
Lately I've been testing out a couple of Dell PowerEdge 2650 machines.
These babies have dual onboard BCM95701A10 NICs (Tigon3 chip) mounted
in the same PCI-X 133MHz 64 bit bus.
Since they have dual onboard GbE, I've been trying to channel bond them
using just two crossover cables between two machines. The results I'm
seeing is at the first glance very strange. What I see is that the
performance when bonded (round robin) is about _half_ (and sometimes even
less) compared to just using a single interface. Here are some netpipe-2.4
results :
64k message size, single interface
1: 65536 bytes 190 times --> 760.54 Mbps in 0.000657 sec
256k message size, single interface
1: 262144 bytes 53 times --> 855.04 Mbps in 0.002339 sec
64 message size, both interfaces (using round robin)
1: 65536 bytes 65 times --> 257.06 Mbps in 0.001945 sec
256k message size, both interfaces (using round robin)
1: 262144 bytes 25 times --> 376.01 Mbps in 0.005319 sec
Looking at the output of netstat -s after a testrun with 256k message
size, I see some differences (main items) :
Single interface :
Tcp:
0 segments retransmited
TcpExt:
109616 packets directly queued to recvmsg prequeue.
52249581 packets directly received from backlog
125694404 packets directly received from prequeue
78 packets header predicted
124999 packets header predicted and directly queued to user
TCPPureAcks: 93
TCPHPAcks: 22981
Bonded interfaces :
Tcp:
234 segments retransmited
TcpExt:
1 delayed acks sent
Quick ack mode was activated 234 times
67087 packets directly queued to recvmsg prequeue.
6058227 packets directly received from backlog
13276665 packets directly received from prequeue
6232 packets header predicted
4625 packets header predicted and directly queued to user
TCPPureAcks: 25708
TCPHPAcks: 4456
The biggest difference as far as I can see is the 'packtes header
predicted', 'packets header predicted and directly queued to user',
'TCPPureAcks' and TCPHPAcks.
I have an idea that this happens because the packets are comming out of
order into the receiving node (i.e the bonding device is alternating
between each interface when sending, and when the receiving node gets the
packets it is possible that the first interface get packets number 0, 2,
4 and 6 in one interrupt and queues it to the network stack before packet
1, 3, 5 is handled on the other interface).
If this is the case, any ideas how to fix this...
I would really love to get 2Gbit/sec on these machines....
PS
I've also seen this feature on the Intel GbE cards (e1000), but these
drivers has a parameter named RxIntDelay which can be set to 0 to get
interrupt for each packet. Is this possible with the tg3 driver too ?
DS
Regards,
--
Steffen Persvold | Scali AS
mailto:[email protected] | http://www.scali.com
Tel: (+47) 2262 8950 | Olaf Helsets vei 6
Fax: (+47) 2262 8951 | N0621 Oslo, NORWAY
On Tue, Aug 27, 2002 at 08:22:03PM +0200, Steffen Persvold wrote:
> I have an idea that this happens because the packets are comming out of
> order into the receiving node (i.e the bonding device is alternating
> between each interface when sending, and when the receiving node gets the
> packets it is possible that the first interface get packets number 0, 2,
> 4 and 6 in one interrupt and queues it to the network stack before packet
> 1, 3, 5 is handled on the other interface).
You pointed your finger on this exact common problem.
You can use the XOR bonding mode (modprobe bonding mode=2), which uses a
hash of mac addresses to select the outgoing interface. This is interesting
if you have lots of L2 hosts on the same network switch.
Or if you have a few hosts on the same switch, you'd better use the "nexthop"
parameter of "ip route". IIRC, it should be something like :
ip route add <destination> nexthop dev eth0 nexthop dev eth1
but read the help, I'm not certain.
Cheers,
Willy
Sorry for reposting this one guys, but I noticed that my original email
had no subject (which in some cases doesn't get peoples attention :)
On Tue, 27 Aug 2002, Steffen Persvold wrote:
> Dear list people,
>
> Lately I've been testing out a couple of Dell PowerEdge 2650 machines.
> These babies have dual onboard BCM95701A10 NICs (Tigon3 chip) mounted
> in the same PCI-X 133MHz 64 bit bus.
>
> Since they have dual onboard GbE, I've been trying to channel bond them
> using just two crossover cables between two machines. The results I'm
> seeing is at the first glance very strange. What I see is that the
> performance when bonded (round robin) is about _half_ (and sometimes even
> less) compared to just using a single interface. Here are some netpipe-2.4
> results :
>
> 64k message size, single interface
> 1: 65536 bytes 190 times --> 760.54 Mbps in 0.000657 sec
>
> 256k message size, single interface
> 1: 262144 bytes 53 times --> 855.04 Mbps in 0.002339 sec
>
> 64 message size, both interfaces (using round robin)
> 1: 65536 bytes 65 times --> 257.06 Mbps in 0.001945 sec
>
> 256k message size, both interfaces (using round robin)
> 1: 262144 bytes 25 times --> 376.01 Mbps in 0.005319 sec
>
> Looking at the output of netstat -s after a testrun with 256k message
> size, I see some differences (main items) :
>
> Single interface :
> Tcp:
> 0 segments retransmited
>
> TcpExt:
> 109616 packets directly queued to recvmsg prequeue.
> 52249581 packets directly received from backlog
> 125694404 packets directly received from prequeue
> 78 packets header predicted
> 124999 packets header predicted and directly queued to user
> TCPPureAcks: 93
> TCPHPAcks: 22981
>
>
> Bonded interfaces :
> Tcp:
> 234 segments retransmited
>
> TcpExt:
> 1 delayed acks sent
> Quick ack mode was activated 234 times
> 67087 packets directly queued to recvmsg prequeue.
> 6058227 packets directly received from backlog
> 13276665 packets directly received from prequeue
> 6232 packets header predicted
> 4625 packets header predicted and directly queued to user
> TCPPureAcks: 25708
> TCPHPAcks: 4456
>
>
> The biggest difference as far as I can see is the 'packtes header
> predicted', 'packets header predicted and directly queued to user',
> 'TCPPureAcks' and TCPHPAcks.
>
> I have an idea that this happens because the packets are comming out of
> order into the receiving node (i.e the bonding device is alternating
> between each interface when sending, and when the receiving node gets the
> packets it is possible that the first interface get packets number 0, 2,
> 4 and 6 in one interrupt and queues it to the network stack before packet
> 1, 3, 5 is handled on the other interface).
>
> If this is the case, any ideas how to fix this...
>
> I would really love to get 2Gbit/sec on these machines....
>
>
> PS
>
> I've also seen this feature on the Intel GbE cards (e1000), but these
> drivers has a parameter named RxIntDelay which can be set to 0 to get
> interrupt for each packet. Is this possible with the tg3 driver too ?
>
> DS
>
> Regards,
--
Steffen Persvold | Scali AS
mailto:[email protected] | http://www.scali.com
Tel: (+47) 2262 8950 | Olaf Helsets vei 6
Fax: (+47) 2262 8951 | N0621 Oslo, NORWAY
From: Steffen Persvold <[email protected]>
Date: Wed, 28 Aug 2002 10:06:19 +0200 (CEST)
> I have an idea that this happens because the packets are comming out of
> order into the receiving node (i.e the bonding device is alternating
> between each interface when sending, and when the receiving node gets the
> packets it is possible that the first interface get packets number 0, 2,
> 4 and 6 in one interrupt and queues it to the network stack before packet
> 1, 3, 5 is handled on the other interface).
That is exactly what is happening. Packets are being reordered.
Welcome to one of the flaws of round-robin trunking. :-)
> If this is the case, any ideas how to fix this...
Don't use round-robin, choose the output device based upon
hashing of some bits in the IP/TCP headers :-)
You won't get 2Gb/sec for a single TCP stream, but you will
for 2 or more.