Hello!
I spend some time investigating why switching runtime* tasks to real-time scheduling policies increases response time dispersion, while the opposite is expected. The main reason is hyper-threading. rt-scheduler tries only to load all logical CPUs, selecting topologically closest when the current is busy. If hyper-threading is enabled, this strategy is counter-productive: tasks are suffering on busy HT-threads when there is a plenty of idle physical cores. Also, rt-scheduler doesn't try to balance rt load between physical CPUs. It's significant because of turbo-boost and frequency scaling technologies: per-core performance depends on the number of idle cores in the same physical cpu. Are there any known solutions of this problem except disabling hyper-threading and frequency scaling at all? Are there any common plans to enhance the load balancing algorithm in the rt-scheduler? Does anyone use rt-scheduler for runtime-like cpu-bound tasks? Why just don't use CFS? :-) Rt-scheduler with modified load balancing shows much better results. I have a prototype (still incomplete and with many dirty hacks), that shows 10-15% performance increase in our production. (*) A simplified model can be described as following: there is one process per machine, with one thread, that receives request from network and puts them into queue; n (n ~ NCPU + 1) worker threads, that get requests from the queue and handle them. Load is cpu-bound, tens of milliseconds per request. Typical CPU load is between 40% and 70%. A typical system has two physical x86-64 cpus with 8-16 physical cores each (x2 with hyper-threading). Thanks, Roman -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
