So I'm trying to get a VERY high throughput benchmark going and we're eating through bandwidth like it's going out of style on our LAN. Trying to figure out a network topology that maximizes the number of Gb links into my cluster so the load drivers have as much room as possible, but I'm limited a bit by the hardware I'm using since it's all extremely low footprint stuff physically. A couple questions:
1) I'm load balancing using my own BIND server and dns round robin. Is there any way to serve up an certain ip address or addresses based on the ip of the requesting host? I.e., look at the originating ip of the query and based on some rule respond with a certain ip address? The trick is that all of the client machines have to hit the same bind server, so I can't solve this by having subsets of the driver machines just point to different dns servers or look up different host names.. they have to use the same dns server and must all look up the same name. I've poked around online but haven't found much on this.. it might be one level higher than my current understanding of zones and BIND configuration.
2) Excuse my ignorance of the nuances of network load balancing as I'm primarily a software guy, but what's the precedent for combining links between two switches to act as a single link? I.e. connect two switches with 4 cables and get a 'simulated' 4Gb link? Ideally just load balancing the packets amongst the links would be fine, but I'm assuming there's some special configurations that would be needed that are likely switch dependent. Are there standards for this sort of thing?
I can't really go into too much detail with respect to the topology as that might give too much away regarding what we're doing, but I'm going to look into it. I really think I have the right building blocks here but need to figure out how to get it all straight and still fulfill the rules of the game.
Thanks in advance for any ideas.
1) I'm load balancing using my own BIND server and dns round robin. Is there any way to serve up an certain ip address or addresses based on the ip of the requesting host? I.e., look at the originating ip of the query and based on some rule respond with a certain ip address? The trick is that all of the client machines have to hit the same bind server, so I can't solve this by having subsets of the driver machines just point to different dns servers or look up different host names.. they have to use the same dns server and must all look up the same name. I've poked around online but haven't found much on this.. it might be one level higher than my current understanding of zones and BIND configuration.
2) Excuse my ignorance of the nuances of network load balancing as I'm primarily a software guy, but what's the precedent for combining links between two switches to act as a single link? I.e. connect two switches with 4 cables and get a 'simulated' 4Gb link? Ideally just load balancing the packets amongst the links would be fine, but I'm assuming there's some special configurations that would be needed that are likely switch dependent. Are there standards for this sort of thing?
I can't really go into too much detail with respect to the topology as that might give too much away regarding what we're doing, but I'm going to look into it. I really think I have the right building blocks here but need to figure out how to get it all straight and still fulfill the rules of the game.
Thanks in advance for any ideas.
