On Fri, Aug 11, 2017 at 02:11:32PM +0200, Lukas Czerner wrote: > Thanks, for the answer. I do not know too much about the DAX enabled HW. > However I do know that there is some variety to it, some can be faster > than DRAM, some can be slower, or on-par with DRAM. Some can be more > expensive, hence probably smaller, some cheaper and bigger. What about > latency and throughput can there be difference as well ? Or it could be similar for reads and much slower for writes. Or it might prefer larger access sizes than a cache line. > That said, it seems to me that there can be some user choice involved in > this at least based on the fact that when DAX is used system memory is not > used. All of the above are charateristics of the medium, not of the application. > So for example when DAX HW is slower than system memory, user can make > a choice to exclude some inodes to speed up particular workload, while > saving system memory where it does not matter as much. What should be factored into the decision might be access hints from the applications, which would be things like madvise/fadvise hints. But the application should not even have to know about something like DAX, and the detailed access characterisics of the medium. And even more importantly we should not encode those complex characterisitcs (see above) into a binary flag in the on-disk format.
