Efficient static parameters about fastclick HOT 7 CLOSED

Well I'm happy to say this feature will be (partially) the subject of our upcoming ASPLOS'21 paper "PacketMill". Therefore closing it :) See https://github.com/aliireza/packetmill

from fastclick.

Maybe @cffs you have input to propose using C++ 14+ language facilities?
Again we could use source-to-source rewriting.

from fastclick.

If we manage to have a tool to propagate configuration parameters into the element code, and generate a number of element variants for different parameter sets, there is a risk of code size explosion, which might actually be detrimental to overall performance.
Maybe it should be restricted to a few important parameters which really affect the fastpath performance. Those could be chosen by the element developer (e.g. using the template approach outlined above), or by the configuration author (by adding a configuration keyword, something like FromDPDKDevice(STATIC true, ...), and then calling some click-propagate-params tool).

from fastclick.

Do you thing code size is really a problem? The unused code will not be accessed, it will sit in RAM, yes, but RAM is cheap and as long as it does not pollute the cache we do not really care.
High-speed equipments-wise. We should definitely keep a "small binary" option that deactivates the feature, as some people use Click in embedded devices.

from fastclick.

I was more thinking about instructions cache problems.
If the different variants are served by different cores (e.g. multiple FromDPDKDevice on different threads), this is not an issue, and specializing would likely be better.
However, you might also have many variants of an element on the same path and served by the same core. In this case, it may happen that instructions from one variant need to be trashed from cache to make room for instructions of another variant (depending on cache size and amount of code on the path). That can be costly if done for every batch of packets. In this case, a single element with one additional branch might give better performance if it can avoid caching issues.

from fastclick.

Ok we agree then. Let's allow specific designated parameters, and somehow warn the user to keep in mind not to use specialized versions too much. The template idea force the specialization though...

from fastclick.

a Big Congrats!

from fastclick.