On Thu, Dec 05, 2013 at 09:39:03AM -0500, Tejun Heo wrote: > Hello, > > On Thu, Dec 05, 2013 at 03:37:45PM +0100, Oleg Nesterov wrote: > > > I'm adding Christophe in Cc because he is interested in tweaking the > > > affinity of call_usermodehelper for cpu isolation. This welcome > > > cleanup confirms that we want to take the direction of being able to > > > change the affinity of workqueue themselves instead of just > > > call_usermodehelper() alone. > > > > OK, but I'd like to remind just in case, as Tejun pointed out this > > patch is wrong ;) > > > > And "change the affinity of workqueue themselves" is not simple, but > > we can make khelper_wq WQ_SYSFS. > > Maybe workqueue should implement and expose default attributes which > are inherited by all workqueues unless they're explicitly overridden? > The use case here is not really about isolating certain subgroup of > workers but rather being able to control the default behavior, right?
So the need I know of is to perform CPU isolation on a given set of CPUs. Their are two known usecases: 1) Optimize the use of the CPU and avoid being disturbed with interrupts, other tasks, etc... These disturbances trash the cache and steal cputime. We want to maximize the use of the CPU. So that's the HPC case. 2) Guarantee that the CPU won't be disturbed while some high priority tasks are running. So we have a critical task that has something to do under a given time deadline or some bounded latency. That's the real time case. Case 1) can deal with a few rare noise, but case 2) can't. In both cases I think that people are happy with having one CPU that handles all the housekeeping: handle timekeeping, unbound timers, workqueues, RCU callbacks and random stats maintainance. But this housekeeping could possibly be divided in the future with some per NUMA node affine housekeepers, who knows? So I think it may be nice to have some flexibility in how we affine these workqueues, hence a user-driven action taken through exisiting interfaces like sysfs sounds like a good candidate. > Thanks. > > -- > tejun -- 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/
