DNS load balancing binds multiple IPs to a domain name. It is good for high bandwidth, but is hard to change dynamically.
Router balancing of connection is similar to the scenario described in this slide, where a router or a load balancer makes decision for the whole connection.
Router balancing of requests in a connection is a more fine-grained load balance method, since it makes separate decisions for the requests within a connection.
pagerank
DNS is a way to load balancing by nature because you can map one domain name to multiple IP addresses. It is out of band, so it can distribute arbitrary bandwidth. But it needs to update all levels of DNS servers when changes happen. This update can take days and is not easy to control.
POTUS
I feel like I'm having a hard time differentiating between strategies to improve latency and throughput. Doesn't scaling out decrease latency too, for the same reasons as the multi-threaded 213 proxy server?
hweetvpu
@POTUS I think latency can be due to various reasons... queueing, computation, disk load, network, etc. In this case scaling out probably helps with queueing latency, but not so much with the latency required for DB load.
hzxa21
@POTUS Latency is the time between client issues a request and the client receives a response. As long as there are no pending requests in the queue just as @hweetvpu said, the latency will be the same regardless of whether you scale out.
vadtani
what @hzxa21 says actually makes sense, scaling out does not decrease latency but can help avoid the increase in latency.
Another commonly used term for scale out is Horizontal Scaling
Several mechanisms for load balancing:
DNS load balancing binds multiple IPs to a domain name. It is good for high bandwidth, but is hard to change dynamically.
Router balancing of connection is similar to the scenario described in this slide, where a router or a load balancer makes decision for the whole connection.
Router balancing of requests in a connection is a more fine-grained load balance method, since it makes separate decisions for the requests within a connection.
DNS is a way to load balancing by nature because you can map one domain name to multiple IP addresses. It is out of band, so it can distribute arbitrary bandwidth. But it needs to update all levels of DNS servers when changes happen. This update can take days and is not easy to control.
I feel like I'm having a hard time differentiating between strategies to improve latency and throughput. Doesn't scaling out decrease latency too, for the same reasons as the multi-threaded 213 proxy server?
@POTUS I think latency can be due to various reasons... queueing, computation, disk load, network, etc. In this case scaling out probably helps with queueing latency, but not so much with the latency required for DB load.
@POTUS Latency is the time between client issues a request and the client receives a response. As long as there are no pending requests in the queue just as @hweetvpu said, the latency will be the same regardless of whether you scale out.
what @hzxa21 says actually makes sense, scaling out does not decrease latency but can help avoid the increase in latency.