| |||||||||||||
O_HOT and O_COLDO_HOT and O_COLDPosted Apr 26, 2012 8:40 UTC (Thu) by neilbrown (subscriber, #359)Parent article: O_HOT and O_COLD
This sounds just like the LBA hinting mentioned briefly here: https://lwn.net/Articles/489311/
The idea is to tell your disk drive how you might expect to use the data later, and it will just sort everything out for you. Promise.
Maybe we should just encode the T10 spec in the VFS and allow filesystems to accept exactly the same hints and do the same things. Or maybe not.
I think if you really need that sort of control, you should make it all much more explicit. Create a /hot filesystem and a /cold filesystem and let that be that.
(Log in to post comments)
O_HOT and O_COLD Posted Apr 26, 2012 19:35 UTC (Thu) by ejr (subscriber, #51652) [Link]
And then a unionfs binding of /hot and /cold, along with userspace tools based on common machine learning techniques that manage /hot and /cold.
The question becomes what are the sufficient statistics necessary for control of what file goes where, and how (and how often) does the kernel update those statistics. While that *could* be handled at user level, intercepting open/close/read/write calls is kludgy and inefficient. IMHO, gathering the right statistics is where the kernel can help.
|
|||||||||||||