25.04.2014, 19:11, "Peter Zijlstra" <pet...@infradead.org>: > On Fri, Apr 25, 2014 at 07:02:16PM +0400, Roman Gushchin wrote: > >> Hm. What I really want (and try to implement), is >> "work as if ht is disabled if there are free physical cores, start using ht >> siblings otherwise". > > At which point I have to ask, what about the rest of the topology?
My prototype works as follows: all physical cores on every numa node are linked into circular list. When I have to select a cpu, I traverse the list and search for a free core. If there is one, I select them. Otherwise, I jump to the other node and search there too. It's better to save symmetry here: when I start with local core number 3, it's reasonable to start with remote core number 3 too. Also, there is a "node balancing" logic, that causes me to start searching with remote node, if a big imbalance is detected. It helps to work good under small load (<8 requests). When there are no free cores, I use the similar logic to find a free thread. Now I'm trying to make this algorithm more general and scalable. It will be great to use the current O(1) cpupri approach on each cpu topology level somehow, but I have no complete solution yet. > Also, how is a task to know if its the 16th or 17th and thus should > expect worse latency? No way. >> It's a 32-thread processor with 16 physical cores. > > No NUMA? I'm not aware of single node systems with 16 cores. Of course, 2 physical processors with 8 cores each :) Sorry. 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/
