Packet copies: Expensive or cheap?

Re:Packet copies: Expensive or cheap? by abelsson on Sunday November 01, 2015 @02:01AM (#10485686)

If you think the cost of the instructions to express copying is the main problem I'd call your mental model flawed. That is not the thing to worry about, when moving data the issued instructions matters very little.

No, the thing to worry about is the memory subsystem. How does the data interact with the caches? What DRAM access patterns do I have? How do I flow data as efficiently as possible?

[ Parent | on HN ]

Re:Packet copies: Expensive or cheap? by nl on Sunday November 01, 2015 @04:43AM (#10485948)

Not if you are writing a software based switch, especially one designed to be scriptable.

Which the author is.

Context is everything.

[ Parent | on HN ]

Re:Packet copies: Expensive or cheap? by wmf on Sunday November 01, 2015 @07:43AM (#10486247)

Can you elaborate? Is a a software based switch exempt from the memory subsystem?

[ Parent | on HN ]

Re:Packet copies: Expensive or cheap? by pmelendez on Sunday November 01, 2015 @08:46AM (#10486374)

I'm not OC but reading the author's project list, it seems to me that he is working heavily in networking optimization in cases where packages are small enough that won't be invalidated from cache until they are done.

If that were true, then the memory subsystem has little to play in this. However, users of this still have to take the memory in consideration when sending packages, hence the OC's comment about context is everything.

[ Parent | on HN ]