In the ongoing 8.4. efforts we’re currently testing the effects of using early write submits – both to local storage and to the peer node(s).
Ie., when DRBD can guess in advance that write requests will soon be sent, it can prematurely send the data pages to the other node; so if the application then does write to storage, all that is needed is a small “
do it” packet. The smaller packet size can be transmitted over the (much faster) meta-data DRBD connection, and so reduces latency by a fair amount.
See this performance improvements in a trial run:
For configuration there’s a new item in the
early-write, using a time value in the usual tenths-of-a-second unit. Eg.
10 will cause DRBD to send the data one second before the application tries to write it.
You can expect that feature in the next proprietary 8.4.5 release of DRBD, so stay tuned!