[Beowulf] Bonding on the Gigabit Ethernet

Ilya ilya at gray-world.net
Fri Jul 8 00:04:16 EDT 2005


Yes, I’ve played with that when I realized that packets has arrive out of order. But it doesn’t
help us much more :

1) additional consuming of processor time for sorting  packages in the buffer on the recipient side
2) “сcongestion window”. This parameter is defined how many packets can be send without ACK. So, it
has an influence on the overall throughput of the TCP stream. And while stream has no DUP ACK (or
losses) congestion window is increasing. But when we received DUP ACK, congestion window is
decreasing, even without retransmission.

Even artificial blocking of DUP ACK doesn’t help.
>From paper : “Experiments have shown, that at artificial blocking the duplicated
acknowledgement may raise productivity of bonding on two Gigabit interfaces from 0.9 Gb/sec up to
1.3 Gb/sec. Expected 1.8 Gb/sec was not reached  due to big time spent on sorting of packages in
the buffer on the recipient side. Thus, for the system containing 2 processors
by 2.4 GHz everyone, throughput of a network with two Gigabit interfaces and use standard bonding
makes a maximum 1.3 Gb/sec.” That is why it is so necessary to achieve that recipient side can
recieve TCP packages by way in turn.

7/7/2005, "Tim Mattox" <tmattox at gmail.com> вы писали:

>Thank you posting your paper.
>
>If you still have your testing rig available, I would be interested to see
>what happens to the performance numbers as you vary the value of
>
>/proc/sys/net/ipv4/tcp_reordering
>
>The node will presume that a packet has been lost if it
>sees "tcp_reording" number of packets out of order.
>tcp_reordering defaults to 3, and can go as high as 127.
>At least this is my understanding of how this sysctl works.  I haven't
>had the appropriate test rig (or time) to really find out.
>
>On 7/7/05, Ilya <ilya at gray-world.net> wrote:
>> Hi all!
>>
>> A little paper about Bonding on the Gigabit Ethernet :
>>
>> http://tom.imm.uran.ru/~u1330/bonding.pdf
>>
>> ---------------------------------------
>> Audi, Vide, Tace, si vis vivere in pace
>>
>> _______________________________________________
>> Beowulf mailing list, Beowulf at beowulf.org
>> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
>
>--
>Tim Mattox - tmattox at gmail.com
>  http://homepage.mac.com/tmattox/
>    I'm a bright... http://www.the-brights.net/
>
>_______________________________________________
>Beowulf mailing list, Beowulf at beowulf.org
>To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
>

---------------------------------------
Audi, Vide, Tace, si vis vivere in pace

_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf



More information about the Beowulf mailing list