Answer the question
In order to leave comments, you need to log in
Has anyone developed a new LAN model for a cluster?
It is necessary to develop a new model of a local area network for high-performance clusters for a diploma.
It is necessary to make a model with better throughput / less redundancy compared to existing analogues, based on the 100g ethernet standard and the diploma of my predecessor, who already made some version of the new model based on the classic OSI model (its main changes are truncated protocol headers ).
I was advised to pay attention to routing, to make it more flexible, or something.
What exactly can be done with it?
In what other direction can you dig?
Answer the question
In order to leave comments, you need to log in
To get started, open rfc over ethernet, draw, analyze, think about what protocols can be sacrificed, perhaps you can sacrifice not only headers, but entire protocols or osi levels, if instead of flexible mechanisms, you make a hard-configured network configuration. A diploma is not so much an implementation as a research work. Good luck to you.
Need to make a modelIn what environment is modeling and testing expected?
some variant of a new model based on the classic OSI modelWhat is the "classic OSI model"? If there is a classical one, then there is also a modernist one, or did I not understand something?
I was advised to pay attention to routing, to make it more flexible, or something.If I were you, I would pay attention to SDN (Software-defined networking, stylish, fashionable, youthful). One controller communicates with several routers (not only in the L3 sense) and each sets specific entries in the routing/forwarding table. From pluses - flexibility, programmability, absence of problems like "ships in the night". I would even suggest that it is most logical to use SDN in data center networks (compared to, say, WAN networks). But there are downsides as well.
In what other direction can you dig?Clos networks (I'm not sure if this is the correct term, the English name is Clos network/fabric). Setting a constant frame size (easier to predict buffer utilization). Refusal to calculate the checksum (transfer of this responsibility to the higher levels of the "classical OSI model"). Rejection of STP (either by adding a hop count field to the ethernet frame, or by assigning unequal roles in terms of forwarding to switch ports, ports from one category can only forward traffic to ports of another category)
Didn't find what you were looking for?
Ask your questionAsk a Question
731 491 924 answers to any question