Quoting Matthew Brost (2021-01-28 22:56:04) > On Mon, Jan 25, 2021 at 02:01:15PM +0000, Chris Wilson wrote: > > Replace the priolist rbtree with a skiplist. The crucial difference is > > that walking and removing the first element of a skiplist is O(1), but > > O(lgN) for an rbtree, as we need to rebalance on remove. This is a > > hindrance for submission latency as it occurs between picking a request > > for the priolist and submitting it to hardware, as well effectively > > trippling the number of O(lgN) operations required under the irqoff lock. > > This is critical to reducing the latency jitter with multiple clients. > > > > The downsides to skiplists are that lookup/insertion is only > > probablistically O(lgN) and there is a significant memory penalty to > > as each skip node is larger than the rbtree equivalent. Furthermore, we > > don't use dynamic arrays for the skiplist, so the allocation is fixed, > > and imposes an upper bound on the scalability wrt to the number of > > inflight requests. > > > > This is a fun data structure but IMO might be overkill to maintain this > code in the i915. The UMDs have effectively agreed to use only 3 levels, > is O(lgN) where N == 3 really a big deal? With GuC submission we will > statically map all user levels into 3 buckets. If we are doing that, do > we even need a complex data structure? i.e. Could use just use can > array of linked lists? Because we need to scale the bst to handle a unqiue key per request with thousands of requests [this is not only about priorities]. And as you will see from the results, even with just a single priority in the system (so one entry in either the skiplist or rbtree), the skiplist is beating the rbtree as measured by the lock hold time around insert/dequeue of requests. That surprised me. -Chris _______________________________________________ Intel-gfx mailing list Intel-gfx@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/intel-gfx
