As discussed in the ML [1], there are some issues around the way ram.c accounts for transferred data and dirty pages for the auto convergence calculations. The patches in these series fix that by:
- Keeping bytes_xfer_prev zeroed until it's used - Ensure bytes_xfer_* counters are accounted for the same period as dirty_pages_* counters - Start throttling in a consistent manner [1] http://lists.nongnu.org/archive/html/qemu-devel/2017-05/msg05524.html Felipe Franciosi (4): migration: keep bytes_xfer_prev init'd to zero migration: set dirty_pages_rate before autoconverge logic migration: set bytes_xfer_* outside of autoconverge logic migration: use dirty_rate_high_cnt more aggressively migration/ram.c | 23 +++++++++++------------ 1 file changed, 11 insertions(+), 12 deletions(-) -- 1.9.5