On Tuesday 07 March 2006 22:19, Benjamin LaHaise wrote: > At this point I'd just like to stir up some discussion, so please comment > away with any ideas and concerns.
I think the first step for better spreading of work would be better MSI-X support. In particular we need ways for drivers to easily set IRQ affinity to CPUs. Then the network drivers could do per CPU TX completion interrupts. This would be relatively easy without too much stack changes. It's on my todo list to implement the interrupt infrastructure for that. Hopefully soon. This should be HT and dual core aware of course. Also we need to teach user space irqbalanced to keep its fingers away from these bound interrupts. [I know the following is in contrary to DaveM's somewhat vague net channel ideas, but it's much easier and less risky:] Later devices that support multiple RX queues with hashing could use the same infrastructure to distribute the receive interrupt and therefore softirq to different CPUs. Now this would still require a CPU<->CPU transfer because the target CPU selected by the hash is unlikely to be the CPU where the receive process runs. One way to fix that would be to check if the scheduler wakeup path could be gotten to slowly migrate the task to the right CPU to avoid the cache line transfers from the interrupting CPU to the final user. I don't know if it would work out - my experiences in writing a NUMA scheduler that tries similar things made me a bit sceptical - but perhaps it would be worth another try. And alternative way would be to use netchannel ideas to make the CPU<-> CPU transfer more cheap. In particular I liked the concept of using arrays of pointers as queues instead of double linked lists to be more cache friendly. -Andi - To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html
