> From: Thomas Gleixner [mailto:t...@linutronix.de] > Sent: Wednesday, November 18, 2015 10:25 AM > Folks! > > After rereading the mail flood on CAT and staring into the SDM for a while, I > think we all should sit back and look at it from scratch again w/o our > preconceptions - I certainly had to put my own away. > > Let's look at the properties of CAT again: > > - It's a per socket facility > > - CAT slots can be associated to external hardware. This > association is per socket as well, so different sockets can have > different behaviour. I missed that detail when staring the first > time, thanks for the pointer! > > - The association ifself is per cpu. The COS selection happens on a > CPU while the set of masks which are selected via COS are shared > by all CPUs on a socket. > > There are restrictions which CAT imposes in terms of configurability: > > - The bits which select a cache partition need to be consecutive > > - The number of possible cache association masks is limited > > Let's look at the configurations (CDP omitted and size restricted) > > Default: 1 1 1 1 1 1 1 1 > 1 1 1 1 1 1 1 1 > 1 1 1 1 1 1 1 1 > 1 1 1 1 1 1 1 1 > > Shared: 1 1 1 1 1 1 1 1 > 0 0 1 1 1 1 1 1 > 0 0 0 0 1 1 1 1 > 0 0 0 0 0 0 1 1 > > Isolated: 1 1 1 1 0 0 0 0 > 0 0 0 0 1 1 0 0 > 0 0 0 0 0 0 1 0 > 0 0 0 0 0 0 0 1 > > Or any combination thereof. Surely some combinations will not make any > sense, but we really should not make any restrictions on the stupidity of a > sysadmin. The worst outcome might be L3 disabled for everything, so what? > > Now that gets even more convoluted if CDP comes into play and we really > need to look at CDP right now. We might end up with something which looks > like this: > > 1 1 1 1 0 0 0 0 Code > 1 1 1 1 0 0 0 0 Data > 0 0 0 0 0 0 1 0 Code > 0 0 0 0 1 1 0 0 Data > 0 0 0 0 0 0 0 1 Code > 0 0 0 0 1 1 0 0 Data > or > 0 0 0 0 0 0 0 1 Code > 0 0 0 0 1 1 0 0 Data > 0 0 0 0 0 0 0 1 Code > 0 0 0 0 0 1 1 0 Data > > Let's look at partitioning itself. We have two options: > > 1) Per task partitioning > > 2) Per CPU partitioning > > So far we only talked about #1, but I think that #2 has a value as well. Let > me > give you a simple example. > > Assume that you have isolated a CPU and run your important task on it. You > give that task a slice of cache. Now that task needs kernel services which run > in kernel threads on that CPU. We really don't want to (and cannot) hunt > down random kernel threads (think cpu bound worker threads, softirq > threads ....) and give them another slice of cache. What we really want is: > > 1 1 1 1 0 0 0 0 <- Default cache > 0 0 0 0 1 1 1 0 <- Cache for important task > 0 0 0 0 0 0 0 1 <- Cache for CPU of important task > > It would even be sufficient for particular use cases to just associate a > piece of > cache to a given CPU and do not bother with tasks at all. > > We really need to make this as configurable as possible from userspace > without imposing random restrictions to it. I played around with it on my new > intel toy and the restriction to 16 COS ids (that's 8 with CDP > enabled) makes it really useless if we force the ids to have the same meaning > on all sockets and restrict it to per task partitioning. > > Even if next generation systems will have more COS ids available, there are > not going to be enough to have a system wide consistent view unless we > have COS ids > nr_cpus. > > Aside of that I don't think that a system wide consistent view is useful at > all. > > - If a task migrates between sockets, it's going to suffer anyway. > Real sensitive applications will simply pin tasks on a socket to > avoid that in the first place. If we make the whole thing > configurable enough then the sysadmin can set it up to support > even the nonsensical case of identical cache partitions on all > sockets and let tasks use the corresponding partitions when > migrating. > > - The number of cache slices is going to be limited no matter what, > so one still has to come up with a sensible partitioning scheme. > > - Even if we have enough cos ids the system wide view will not make > the configuration problem any simpler as it remains per socket. > > It's hard. Policies are hard by definition, but this one is harder than most > other policies due to the inherent limitations. > > So now to the interface part. Unfortunately we need to expose this very > close to the hardware implementation as there are really no abstractions > which allow us to express the various bitmap combinations. Any abstraction I > tried to come up with renders that thing completely useless. > > I was not able to identify any existing infrastructure where this really fits > in. I > chose a directory/file based representation. We certainly could do the same
Is this be /sys/devices/system/? Then create qos/cat directory. In the future, other directories may be created e.g. qos/mbm? Thanks. -Fenghua -- 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/
