Do I understand the theory of load on a server with a 100 Mbps channel correctly?

Question

Highload

Greetings! Tell me, please, if I understand correctly the theory of load on a server with a 100 Mbps channel.
Let's say the server only accepts requests, but does not return responses. The server hardware and the executable code on it can withstand processing of 10 thousand / sec. requests.
The amount of data sent to the server is constant and equals 300 bytes in one request, it turns out that the channel can withstand 100 Mbps / 300 bytes = 43690 requests / sec.
Since the hardware and software can only process 10k/sec, is the channel being used inefficiently? In this case, would a channel of about 25 Mbps be enough?
Now, the server only sends responses and is able to process 10k/sec.
The amount of data sent is constant and equals 50 kb in one response, it turns out that the channel can withstand 100 Mbps / 50 kb = 256 responses / sec.
Since the channel can only handle 256 replies/sec, is the hardware being used inefficiently? In this case, would there be enough iron capable of processing 256 requests / sec?
Thank you!

Answer 1

There is logic in your reasoning.
But do not forget:
1. Take into account overhead costs (this includes service traffic and system traffic and delivery confirmations and all handshakes, etc.), which can make up a large% of the target traffic
2. Do not forget about latency. On thick channels, most often, this indicator is lower, since, again, most often, it grows exponentially to the load on the channel, so for a load of 80 megabits for a hundred megabit channel, delays may already appear, for gigabit - definitely not.
3. Always use less than 100% of your power. Not even 80%. It is good practice to use 60% of the capacity at the peak.
4. It is better to have one controlled bottleneck, the parameters of which are known in advance (for example, the channel width), than to have 5 uncontrolled ones, each of which works at different times (either the CPU is in the shelf, then there is not enough memory, then the disk queue, then channel).

Answer 2

When you receive traffic, then approximately 1/10 of its volume goes back to the subscriber in the form of delivery confirmation. That. you have to take that into account too.

Answer 3

So it is, but in reality it is not. rps on different hardware will be different for you. As for traffic, the speed of its incoming / outgoing from outside is also important. For example, the client will send a request to 300b at a speed of 8bit / sec, such a request will live 37.5sec. There are also network costs.
Iron should not be taken with the expectation of a 100% load, this is a road to nowhere.

Answer 4

If you are in judp, then almost yes, since you did not take into account service information. You cannot
work in one direction on tisipi