|
North American Network Operators Group Date Prev | Date Next | Date Index | Thread Index | Author Index | Historical Re: customers and web servers and level one naps
> Easy does it Stephen..... sorry I misunderstood you. I have not yet begun > to work on the tape. So let me reorient my question. > > You encourage ISPs to place web servers ON THEIR OWN NETWORKS, BEHIND > THEIR ROUTERS. > > I see now the point you are making and it is a critical one, but please > have mercy when i make a mistake. Sorry. I've had to defend that point a lot over the past year, and I guess I come out of the blocks pretty fast when I see it nowadays. > Having said this, the web servers are still sited within PAIX and > topologically a lot closer to the exchange switching fabric than they have > been before. This presumably offers some advantages for the preformance > of those machines. The only thing i am trying to ascertain is to whether > this has been tried at other exchanges or not and why. As far as I am > aware it has not. Yes, servers sited at PAIX are topologically much closer to the exchange switching fabric than before. There are two advantages to this: 1. Fewer server responses backhauled over internal intrastructure. Of course, there may be more requests backhauled through internal infrastructure as a result. Grossly characterized, web servers tend to be of the small-request/large-response category, so the net result may be less load on internal infrastructure. Lots of factors come into play, though; who else is peering at the exchange where your server is located, peering policies at all the exchanges where you are present, etc. 2. If, in fact, the server responses wind up having shorter paths to their destinations, with less latency, the time during which server resources are consumed to service a single request goes down. Presumably, throughput goes up. > Bill Manning asked whether PAIX was a major exchange. > > No of course it is not. Although I'd certainly like to think it's going to be. :-) > But bill is your response meant to imply that at > a major exchange, there is simply going to be too much traffic to add the > web stuff? Since the server is BEHIND the customer router the web > traffic would hit the switch as part of the application layer traffic > brought there by the customer. therefor should it really make any > difference to have the web traffic avoid the extra hops of > traversing the local loop? The server is behind the ISP router. Depending on the topology that is chosen by the ISP, the server could be one hop away from the IX switch, rather than two or three with the latency of a wide-area circuit thrown in. Even if the hop count goes up to two or three, if it's all in be same building, these hops could all be some flavor of FDDI, rather than having DS3 circuits to get out of the building. Stephen - - - - - - - - - - - - - - - - -
|