On Tue, Jan 28, 2014 at 2:57 PM, Dan Williams <dan.j.williams@xxxxxxxxx> wrote: > In addition to disk drives with internally tiered-storage (solid-state > + magnetic media) the kernel has also grown native I/O-caching / > tiered-storage implementations in bcache and dm-cache. > > Currently, all these solutions depend on heuristics to determine what > tier the data referenced in an I/O belongs. However, the presence of > hinting proposals from SCSI [1], ATA [2], and bcache [3] indicate that > these devices (hardware or virtual) want to consume explicit hints > indicating the value of caching data in a higher performing tier. > > At LSF I want to discuss options and opportunities for plumbing cache > hints from userspace, through the I/O stack to devices. My colleague, > Jason Akers, is also interested in this discussion as he is > investigating how to exploit these hints from userspace. I > participated in the LSF 2012 discussion of this topic, see that it was > raised again at LSF 2013, and note that we have not settled on an > enabling path. What's new for this year is an effort to set aside, > for now, the deeper complexities of the device specification proposals > and focus on a minimal set of hints that can be specified per-process > (ionice) and maybe per-file (fadvise). > > My suspicion is that AIO attributes is useful for applications that > want access to the full range of access hints exposed by a device. > However, for the general buffered-I/O / tiered-storage case, a small > set of hints achieves the bulk of the value. > > I am also interested in: > Volatile ranges > SMR Enabling > Integrity passthrough > > -- > Dan > > [1] T10 LBA Access Hints > [2] T13 Hybrid Information Feature Correction, the hinting scheme is defined by SATA-IO in SATA 3.2 > [3] AIO Attributes: http://marc.info/?l=linux-aio&m=136580574523674&w=2 -- dm-devel mailing list dm-devel@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/dm-devel
