> 30 апр. 2018 г., в 0:48, Andres Freund <and...@anarazel.de> написал(а): > > On 2018-04-25 11:31:12 +0500, Andrey Borodin wrote: >> >> 1. Teaching BgWriter to used data from eviction strategy to aggressively >> flush data to disk (instead of ++next_to_clean ) >> 2. Implementing strategies as lock-free algorithms for freelist >> These parts seem most important for benchmarking. >> Also: >> 3. Converting all rings to single buffer manager where possible >> 4. Using O_DIRECT while writing data files >> 5. Using aio and scheduling of writes >> These parts are not necessary, but most challenging, while not impossible >> though. > > These largely seem to be increasing the scope beyond reason...
I suspect that performance benefits can be not that big or even marginal, if we do not extend comprehensive eviction strategy with bgwriter fixes and O_DIRECT. But, on the other hand, this suspicion is not based on any real fact. And if eviction strategy is actually good for anything it will show performance benefits on it's own. > 30 апр. 2018 г., в 0:39, Юрий Соколов <funny.fal...@gmail.com> написал(а): > > (btw, why no one mention CLOCK+ in this thread). I do not know what CLOCK+ is. Can you, please, send me a reference. Best regards, Andrey Borodin.
