| Message ID | 20221025044730.319941-1-leobras@redhat.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | MultiFD zero-copy improvements | expand |
On Tue, Oct 25, 2022 at 01:47:27AM -0300, Leonardo Bras wrote: > It all works fine in my tests, but maybe I missed some cornercase. > Please provide any feedback you find fit. Personally I very much like the whole series. :) I just want to make sure it won't start to fail users when switching to this new version of qemu due to vmlocked restrictions which I expect to grow. But I see that there're discussions on patch 4 for further optimizations of lockvm and let's keep the discussion there for this. The other thing is since it's a patchset for performance optimization only, it'll always be nice to attach performance numbers. If on slow networks it won't show a difference worth mentioning that. And ultimately I believe anyone would be interested to know in which case (e.g. 100g nic?) would this start to benefit. Thanks!
RFC for an improvement suggested by Juan during the KVM Forum and a few optimizations I found in the way. Patch #1 is just moving code to a helper, should have no impact. Patch #2 is my implementation of Juan's suggestion. I implemented the simplest way I thought on the array size: a fixed defined value. I am not sure if this is fine, or if the array size should be either informed by the user either via QMP or cmdline. That's an important point I really need feedback on. Patch #3: Improve the qio_channel_flush() interface to accept flush waiting for some writes finished instead of all of them. This reduces the waiting time, since most recent writes/sends will take more time to finish, while the older ones are probably finished by the first recvmsg() syscall return. Patch #4 uses #3 in multifd zero-copy. It flushes the LRU half of the header array, allowing more writes to happen while the most recent ones are ongoing, instead of waiting for everything to finish before sending more. It all works fine in my tests, but maybe I missed some cornercase. Please provide any feedback you find fit. Thank you all! Best regards, Leo Leonardo Bras (4): migration/multifd/zero-copy: Create helper function for flushing migration/multifd/zero-copy: Merge header & pages send in a single write QIOChannel: Add max_pending parameter to qio_channel_flush() migration/multifd/zero-copy: Flush only the LRU half of the header array include/io/channel.h | 7 +++- migration/multifd.h | 5 ++- io/channel-socket.c | 5 ++- io/channel.c | 5 ++- migration/multifd.c | 88 ++++++++++++++++++++++++++------------------ 5 files changed, 68 insertions(+), 42 deletions(-)