Specific help of OSI layered model to work

   For example, it is necessary to optimize the performance of complex systems , We can think and practice from different perspectives .

1. In the application layer, not only two indicators should be considered ：TPS( Transactions per second ) and QPS( Query rate per second ), And consider the average size of each message .
2. If you want to improve the performance optimization of handshake or symmetric encryption , It is essentially optimized at the presentation layer , need TLS Security layer to discuss performance optimization .
3. Yes RTT( Network delay ) When optimizing , In essence, we think from the transport layer , You need to consider the throughput of the network card 、 Of the entire host BPS.
4. If you want to TCP When optimizing the performance of reliable transmission , You can think about adding TCP Speed of connection handshake , This is to consider the problem at a very deep optimization level .
5. If you want to PPS( Per second packet) When optimizing , In essence, the problem is considered from the network layer . For example, the back-end or operation and maintenance personnel use Intel DPDK, It is essentially optimized at the network layer .
6. If the data link layer is optimized , So more attention should be paid to broadcasting and ARP The deception of .
7. How to choose load balancing ？ Load balancing consists of three layers 、 four layers 、 Seven layer load balancing . Seven layer load balancing is used to resolve various protocols in the application layer . At the fourth floor, it can only be resolved to UDP perhaps TCP. The three-layer load balancing is the network layer load balancing , use NET And other technologies to balance the burden , For example, common LVS.

   Case a ： Alibaba cloud ECS Storage optimization . In the Alibaba cloud environment , There are many virtual machines running on a physical machine ( Purchased ECS The server , In essence, what you buy is a virtual machine ). If you buy a cloud disk , The cloud disk is automatically attached to the virtual machine . On any virtual machine IO operation , In essence, it will be called through the physical machine TTC The program , Forward it to the remote storage cluster . Before optimization , Storage cluster cache The cache is placed at the remote end . When placed at the far end , How to optimize ？

   Standing on the network layer ( Three layers ) When optimizing from the angle of , The transmission path of the network is too long . There is another problem with long transmission paths , For example, the back-end cloud disk ensures reliability , Will make twoorthree copies , This will cause the cost to be too high . If you move the cache forward , Directly on the physical machine will shorten the path .

   Case 2 ： How to reduce the cost of multi-threaded multiplexing ？ For example, a physical computer runs 20 A virtual machine , On average, each virtual machine contains a system disk and a data disk . Add up to 40 A plate .40 Disks use multithreading to solve problems . But multithreading can cause multiplexing , The cost will be higher . So how to reduce the cost of multiplexing ？ The original multi-threaded architecture is replaced by a co process , In essence, it is optimized at the transport layer .

   Case 2 ： Alibaba cloud uses classic networks , Intranet IP yes 10 At the beginning A Class subnet address , What is used today is B Class subnet address , The network is essentially running in VPC perhaps SDN( Software Defined Network ), This network is on the data link layer or between the data link layer and the network layer (2.5 layer ). It's essentially in rp Message and ip tunnel Conceptual writing .