On 21/05/21 14:58, Srikar Dronamraju wrote: > * Peter Zijlstra <pet...@infradead.org> [2021-05-21 10:14:10]: > >> On Fri, May 21, 2021 at 08:08:02AM +0530, Srikar Dronamraju wrote: >> > * Peter Zijlstra <pet...@infradead.org> [2021-05-20 20:56:31]: >> > >> > > On Thu, May 20, 2021 at 09:14:25PM +0530, Srikar Dronamraju wrote: >> > > > Currently scheduler populates the distance map by looking at distance >> > > > of each node from all other nodes. This should work for most >> > > > architectures and platforms. >> > > > >> > > > However there are some architectures like POWER that may not expose >> > > > the distance of nodes that are not yet onlined because those resources >> > > > are not yet allocated to the OS instance. Such architectures have >> > > > other means to provide valid distance data for the current platform. >> > > > >> > > > For example distance info from numactl from a fully populated 8 node >> > > > system at boot may look like this. >> > > > >> > > > node distances: >> > > > node 0 1 2 3 4 5 6 7 >> > > > 0: 10 20 40 40 40 40 40 40 >> > > > 1: 20 10 40 40 40 40 40 40 >> > > > 2: 40 40 10 20 40 40 40 40 >> > > > 3: 40 40 20 10 40 40 40 40 >> > > > 4: 40 40 40 40 10 20 40 40 >> > > > 5: 40 40 40 40 20 10 40 40 >> > > > 6: 40 40 40 40 40 40 10 20 >> > > > 7: 40 40 40 40 40 40 20 10 >> > > > >> > > > However the same system when only two nodes are online at boot, then >> > > > the >> > > > numa topology will look like >> > > > node distances: >> > > > node 0 1 >> > > > 0: 10 20 >> > > > 1: 20 10 >> > > > >> > > > It may be implementation dependent on what node_distance(0,3) where >> > > > node 0 is online and node 3 is offline. In POWER case, it returns >> > > > LOCAL_DISTANCE(10). Here at boot the scheduler would assume that the >> > > > max >> > > > distance between nodes is 20. However that would not be true. >> > > > >> > > > When Nodes are onlined and CPUs from those nodes are hotplugged, >> > > > the max node distance would be 40. >> > > > >> > > > To handle such scenarios, let scheduler allow architectures to populate >> > > > the distance map. Architectures that like to populate the distance map >> > > > can overload arch_populate_distance_map(). >> > > >> > > Why? Why can't your node_distance() DTRT? The arch interface is >> > > nr_node_ids and node_distance(), I don't see why we need something new >> > > and then replace one special use of it. >> > > >> > > By virtue of you being able to actually implement this new hook, you >> > > supposedly can actually do node_distance() right too. >> > >> > Since for an offline node, arch interface code doesn't have the info. >> > As far as I know/understand, in POWER, unless there is an active memory or >> > CPU that's getting onlined, arch can't fetch the correct node distance. >> > >> > Taking the above example: node 3 is offline, then node_distance of (3,X) >> > where X is anything other than 3, is not reliable. The moment node 3 is >> > onlined, the node distance is reliable. >> > >> > This problem will not happen even on POWER if all the nodes have either >> > memory or CPUs active at the time of boot. >> >> But then how can you implement this new hook? Going by the fact that >> both nr_node_ids and distance_ref_points_depth are fixed, how many >> possible __node_distance() configurations are there left? >> > > distance_ref_point_depth is provided as a different property and is readily > available at boot. The new api will use just use that. So based on the > distance_ref_point_depth, we know all possible node distances for that > platform. > > For an offline node, we don't have that specific nodes distance_lookup_table > array entries. Each array would be of distance_ref_point_depth entries. > Without the distance_lookup_table for an array populated, we will not be > able to tell how far the node is with respect to other nodes. > > We can lookup the correct distance_lookup_table for a node based on memory > or the CPUs attached to that node. Since in an offline node, both of them > would not be around, the distance_lookup_table will have stale values. >
Ok so from your arch you can figure out the *size* of the set of unique distances, but not the individual node_distance(a, b)... That's quite unfortunate. I suppose one way to avoid the hook would be to write some "fake" distance values into your distance_lookup_table[] for offline nodes using your distance_ref_point_depth thing, i.e. ensure an iteration of node_distance(a, b) covers all distance values [1]. You can then keep patch 3 around, and that should roughly be it. >> The example provided above does not suggest there's much room for >> alternatives, and hence for actual need of this new interface. >> > > -- > Thanks and Regards > Srikar Dronamraju
