On Wed, 2009-08-26 at 06:18 +1000, Alex Samad wrote: > On Tue, Aug 25, 2009 at 07:30:08PM +0200, Olle E. Johansson wrote: > > > > 25 aug 2009 kl. 18.50 skrev John A. Sullivan III: > > > > > On Tue, 2009-08-25 at 18:28 +0200, Olle E. Johansson wrote: > > >> 25 aug 2009 kl. 16.20 skrev Olivier: > > [snip] > > > > mode > > > in Linux on any old switch and it works reasonably well other than for > > > some excessive ARP traffic. However, as we found out the hard way > > > when > > > building our Nexenta SAN, bonding works very well with many-to-many > > > traffic but does very little to boost one-to-one network flows. They > > > will all collapse to the same pair of NICs in most scenarios and, in > > > the > > > one mode where they do not, packet sequencing issues will reduce the > > > bandwidth to much less than the sum of the connections. Take care - > > > John > > > > That is very good feedback - thanks, John! > > > > Which means that my plan B needs to be put in action. Well, I did > > create a new branch for it yesterday... ;-) > > any thoughts of different media like 10G ethernet or infiniband ? > > ><snip> Yes, this is drifting a little off-topic but good network design does provide the foundation for good Asterisk design. If we have lots of servers talking to lots of servers, bonding over Gig links works very well. But as we build fewer very big servers via virtualization or, as in this case, trying to make a single large server do the work previously handled by several, the network bandwidth becomes a huge issue. Because almost all bonding algorithms choose a single path for a flow of data (usually based upon MAC address but sometimes on IP address or even socket), bonding becomes less useful in these scenarios. In fact, it is even worse - even in cases where the OS stack (e.g., Linux) will support bonding based upon data above the MAC layer, the switches frequently do not and will again collapse several paths into one as soon as the data crosses the switch.
Thus, for few-to-few traffic patterns, bigger pipes such as 10G are better than bonded pipes. Specifically, 10 bonded 1 Gbps links will effectively yield 1 Gbps throughput as opposed to 1 10Gbps link yielding 10 Gpbs throughput. As an aside, in our iSCSI work, we found latency to be a huge issue if the file block size was small (e.g., Linux files - 4K block size). Thus, the lower latency of faster protocols is a huge performance booster. This will not be so much of an issue with Asterisk where the difference between 100 usecs and 10 usecs in negligible. -- John A. Sullivan III Open Source Development Corporation +1 207-985-7880 jsulli...@opensourcedevel.com http://www.spiritualoutreach.com Making Christianity intelligible to secular society _______________________________________________ -- Bandwidth and Colocation Provided by http://www.api-digital.com -- AstriCon 2009 - October 13 - 15 Phoenix, Arizona Register Now: http://www.astricon.net asterisk-users mailing list To UNSUBSCRIBE or update options visit: http://lists.digium.com/mailman/listinfo/asterisk-users
