Awesome stuff

[FlorianHeigl]

Thank you for not letting this rot away as a student project we'll only ever find papers about - instead you shared it.
And it's feasible to try out, albeit in a smaller scale once one has some pmem enabled systems. Wonderful!

(I wish I were able to port this over to the 48core Cavium stuff, since they also got partitioning. From the paper, you stopped measuring at 8 clients due to line rate exhaustion, as a sysadmin I have to say that is where it starts to get interesting and your parallel "busy" handling looked very good. It would have been interesting to see how it holds up under very high overload, i.e. 1024 clients, as that's the expectation for real enterprise storage to handle gracefully (IOW: an architecture that is load-independent, as to not topple, fall over, implode, fold into itself when queues become excessive. your work seems to have had potential for that, even if it of course was very specific with the use of pmem. Today with nvme direct (gpu direct?) reads it would probably even be able to integrate slower storage in the same manner)

oh well. maybe some lucky day an engineer notices this code and it becomes a thing ;-) till then, again ty for sharing.)