Il 20/02/2013 15:56, Michael S. Tsirkin ha scritto: > > I have tested that the series builds and is bisectable on Linux and Windows > > hosts. But I have not done extensive testing on other host platforms or > > with > > long-term guests to check for performance regressions. > > Wouldn't it make sense to switch to epoll instead? > poll does not scale well to a huge number of descriptors.
At least for the block layer, where we want to go is to distribute the performance-sensitive work to many independent dataplane threads. Then each of them could use epoll (basically a new aio-linux.c file based on aio-posix.c), like hw/dataplane does already IIRC. Only low-priority tasks are left to the iothread. The required work to get there would really be a superset of this, and this series is a very good start. Paolo